Introduction to Hilbert Spaces with Applications, Third Edition

Hilbert Spaces with Applications ´ Lokenath Debnath and Piotr Mikusinski

This page intentionally left blank

Hilbert Spaces with Applications Third Edition Lokenath Debnath University of Texas—Pan American

´ Piotr Mikusinski University of Central Florida

AMSTERDAM • BOSTON • HEIDELBERG • LONDON • NEW YORK • OXFORD • PARIS • SAN DIEGO • SAN FRANCISCO • SINGAPORE • SYDNEY • TOKYO

Elsevier Academic Press 30 Corporate Drive, Suite 400, Burlington, MA 01803, USA 525 B Street, Suite 1900, San Diego, California 92101-4495, USA 84 Theobald’s Road, London WC1X 8RR, UK ∞ This book is printed on acid-free paper.


Copyright © 2005, Elsevier Inc. All rights reserved.

No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Permissions may be sought directly from Elsevier’s Science & Technology Rights Department in Oxford, UK: phone: (+44) 1865 843830, fax: (+44) 1865 853333, E-mail: [email protected]. You may also complete your request on-line via the Elsevier homepage (http://elsevier.com), by selecting “Customer Support” and then “Obtaining Permissions.” Library of Congress Cataloging-in-Publication Data Application Submitted British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library ISBN 13: 978-0-12-2084386 ISBN 10: 0-12-208438-1 For all information on all Elsevier Academic Press publications visit our Web site at www.books.elsevier.com Printed in the United States of America 05 06 07 08 09 10 9 8 7 6 5 4 3 2 1

This book is dedicated to the memory of our fathers: ´ J OGESH C HANDRA D EBNATH and JAN M IKUSI NSKI

This page intentionally left blank

Contents

Preface to the Third Edition xi Preface to the Second Edition xiii Preface to the First Edition xv CHAPTER 1

Normed Vector Spaces 1 1.1 1.2 1.3 1.4 1.5 1.6

Introduction 1 Vector Spaces 2 Normed Spaces 8 Banach Spaces 19 Linear Mappings 25 Contraction Mappings and the Banach Fixed Point Theorem 32 1.7 Exercises 34 CHAPTER 2

The Lebesgue Integral 39 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9

Introduction 39 Step Functions 40 Lebesgue Integrable Functions 45 The Absolute Value of an Integrable Function 48 Series of Integrable Functions 50 Norm in L1 (R) 52 Convergence Almost Everywhere 55 Fundamental Convergence Theorems 58 Locally Integrable Functions 62 vii

viii

Contents

2.10 2.11 2.12 2.13 2.14 2.15 2.16 CHAPTER 3

The Lebesgue Integral and the Riemann Integral 64 Lebesgue Measure on R 67 Complex-Valued Lebesgue Integrable Functions 71 The Spaces Lp (R) 74 Lebesgue Integrable Functions on RN 78 Convolution 82 Exercises 84

Hilbert Spaces and Orthonormal Systems 93 3.1 3.2 3.3 3.4 3.5 3.6 3.7

Introduction 93 Inner Product Spaces 94 Hilbert Spaces 99 Orthogonal and Orthonormal Systems 105 Trigonometric Fourier Series 122 Orthogonal Complements and Projections 127 Linear Functionals and the Riesz Representation Theorem 132 3.8 Exercises 135 CHAPTER 4

Linear Operators on Hilbert Spaces 145 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 4.10 4.11 4.12

CHAPTER 5

Introduction 145 Examples of Operators 146 Bilinear Functionals and Quadratic Forms 151 Adjoint and Self-Adjoint Operators 158 Invertible, Normal, Isometric, and Unitary Operators 163 Positive Operators 168 Projection Operators 175 Compact Operators 180 Eigenvalues and Eigenvectors 186 Spectral Decomposition 196 Unbounded Operators 201 Exercises 211

Applications to Integral and Differential Equations 217 5.1 5.2 5.3 5.4 5.5

Introduction 217 Basic Existence Theorems 218 Fredholm Integral Equations 224 Method of Successive Approximations 226 Volterra Integral Equations 228

Contents

5.6 Method of Solution for a Separable Kernel 233 5.7 Volterra Integral Equations of the First Kind and Abel’s Integral Equation 236 5.8 Ordinary Differential Equations and Differential Operators 239 5.9 Sturm–Liouville Systems 247 5.10 Inverse Differential Operators and Green’s Functions 253 5.11 The Fourier Transform 258 5.12 Applications of the Fourier Transform to Ordinary Differential Equations and Integral Equations 271 5.13 Exercises 279 CHAPTER 6

Generalized Functions and Partial Differential Equations 287 6.1 6.2 6.3 6.4

Introduction 287 Distributions 288 Sobolev Spaces 300 Fundamental Solutions and Green’s Functions for Partial Differential Equations 303 6.5 Weak Solutions of Elliptic Boundary Value Problems 323 6.6 Examples of Applications of the Fourier Transform to Partial Differential Equations 329 6.7 Exercises 343 CHAPTER 7

Mathematical Foundations of Quantum Mechanics 351 7.1 Introduction 351 7.2 Basic Concepts and Equations of Classical Mechanics 352 Poisson’s Brackets in Mechanics 361 7.3 Basic Concepts and Postulates of Quantum Mechanics 363 7.4 The Heisenberg Uncertainty Principle 377 7.5 The Schrödinger Equation of Motion 379 7.6 The Schrödinger Picture 395 7.7 The Heisenberg Picture and the Heisenberg Equation of Motion 401 7.8 The Interaction Picture 405 7.9 The Linear Harmonic Oscillator 407 7.10 Angular Momentum Operators 412 7.11 The Dirac Relativistic Wave Equation 420 7.12 Exercises 423

ix

x CHAPTER 8

Contents

Wavelets and Wavelet Transforms 433 8.1 8.2 8.3 8.4

Brief Historical Remarks 433 Continuous Wavelet Transforms 436 The Discrete Wavelet Transform 444 Multiresolution Analysis and Orthonormal Bases of Wavelets 452 8.5 Examples of Orthonormal Wavelets 462 8.6 Exercises 473 CHAPTER 9

Optimization Problems and Other Miscellaneous Applications 477 9.1 Introduction 477 9.2 The Gateaux and Fréchet Differentials 478 9.3 Optimization Problems and the Euler–Lagrange Equations 490 9.4 Minimization of Quadratic Functionals 505 9.5 Variational Inequalities 507 9.6 Optimal Control Problems for Dynamical Systems 510 9.7 Approximation Theory 517 9.8 The Shannon Sampling Theorem 522 9.9 Linear and Nonlinear Stability 526 9.10 Bifurcation Theory 530 9.11 Exercises 535

Hints and Answers to Selected Exercises 547 Bibliography 565 Index 571

Preface to the Third Edition

The previous two editions of our book were very well received. This new edition preserves the basic content and style of the earlier editions. It is a graduate-level text for students and a research reference for professionals in mathematics, science, and engineering. The theoretical foundations are presented in as simple a way as possible, but without sacrificing the mathematical rigor. In the part devoted to applications, we present a wide variety of topics, from classical applications to some recent developments. While the treatment of those applications is rather brief, our hope is that we present enough to stimulate interest that will encourage readers to further studies in those areas. We have received various comments and suggestions from our colleagues, readers, and graduate students, from the United States and abroad. Those comments have been very helpful in writing this edition. We have made some additions and changes in order to modernize the contents. An effort to improve clarity of presentation and to correct a number of typographical errors was made. New examples and exercises were added. We have also taken the opportunity to entirely rewrite and reorganize several sections in an appropriate manner and to update the bibliography. Some of the major changes and additions include the following: • Chapter 1 has been reorganized, some sections were combined, and the order of material has been modified. • A complete characterization of finite dimensional normed spaces has been added. • A new section on Lp spaces has been added in Chapter 2.

xi

xii

Preface to the Third Edition

• The section on spectral properties of operators in Chapter 4 has been expanded. • The presentation of the Fourier transform has been moved from Chapter 4 to Chapter 5. • A new section on Sobolev spaces has been added to Chapter 6. • Chapter 8 on wavelets and wavelet transforms has been revised and new material added, including a new section on orthonormal wavelets. We would like to take this opportunity to thank all those who helped us improve the book by reading parts of the manuscript and sharing their comments with us, including Andras Balogh, Cezary Ferens, Ziad Musslimani, Zuhair Nashed, and Vladimir Varlamov. In spite of the best efforts of everyone involved, some typographical errors doubtless remain. Special thanks to June Wingler who helped us with the preparation of the LaTeX files. Finally, we wish to express our grateful thanks to Tom Singer, assistant editor, and staff of Elsevier Academic Press for their help and cooperation. Lokenath Debnath, University of Texas — Pan American ´ Piotr Mikusinski, University of Central Florida January 2005

Preface to the Second Edition

When the first edition of this book was published in 1990, it was well received, and we found the comments and criticisms of graduate students and faculty members from the United States and abroad to be helpful, beneficial, and encouraging. This second edition is the result of that input. We have taken advantage of this new edition to update the bibliography and correct typographical errors, to include additional topics, examples, exercises, comments, and observations, and, in some cases, to entirely rewrite whole sections. The most significant difference from the first edition is the inclusion of a completely new chapter on wavelets. We have, however, tried to preserve the character of the first edition. We intend the book to be a source of classical and modern topics dealing fully with the basic ideas and results of Hilbert space theory and functional analysis, and we also intend it to be an introduction to various methods of solution of differential and integral equations. Some of the highlights include the following: • The book offers a detailed and clear explanation of every concept and method that is introduced, accompanied by carefully selected worked examples, with special emphasis being given to those topics in which students experience difficulty. • A wide variety of modern examples of applications has been selected from areas of integral and ordinary differential equations, wavelets, generalized functions and partial differential equations, control theory, quantum mechanics, fluid dynamics and solid mechanics, optimization, calculus of variations, variational inequalities, approximation theory, linear and nonlinear stability analysis, and bifurcation theory.

xiii

xiv

Preface to the Second Edition

• The book is organized with sufficient flexibility to enable instructors to select chapters appropriate to courses of differing lengths, emphases, and levels of difficulty. • A wide spectrum of exercises has been carefully chosen and included at the end of each chapter so the reader may further develop both rigorous skills in the theory and applications of functional analysis and a deeper insight into the subject. Answers and hints to selected exercises are provided at the end of the book to provide additional help to students. It is our pleasure to express our gratitude to those who offered their generous help at different stages of the preparation of this book. Our special thanks are due to Professor Michael Taylor, who read most of the manuscript and suggested many corrections and improvements. Professors Ahmed Zayed and Kit Chan read parts of the manuscript and offered various criticisms and suggestions that have improved the book. June Wingler, with unflagging industry and exemplary patience, typed parts of the manuscript. Finally, we wish to express our grateful thanks to Mr. Charles Glaser, executive editor, and the staff of Academic Press for their help and cooperation. Needless to say, the authors take responsibility for any remaining errors. The final text was typeset using AmSTeX, and the figures were prepared with the aid of Adobe Illustrator 7.0. ´ Lokenath Debnath, Piotr Mikusinski Orlando, January 1998

Preface to the First Edition

Functional analysis is one of the central areas of modern mathematics, and the theory of Hilbert spaces is the core around which functional analysis has developed. Hilbert spaces have a rich geometric structure because they are endowed with an inner product that allows the introduction of the concept of orthogonality of vectors. We believe functional analysis is best approached through a sound knowledge of Hilbert space theory. Our belief led us to prepare an earlier manuscript, which was used as class notes for courses on Hilbert space theory at the University of Central Florida and Georgia Institute of Technology. This book is essentially based on those notes. One of the main impulses for the development of functional analysis was the study of differential and integral equations arising in applied mathematics, mathematical physics, and engineering; it was in this setting that Hilbert space methods arose and achieved their early successes. With ever greater demands for mathematical tools to provide both theory and applications for science and engineering, the utility and interest of functional analysis and Hilbert space theory seems more clearly established than ever. Keeping these things in mind, our main goal in this book has been to provide both a systematic exposition of the basic ideas and results of Hilbert space theory and functional analysis, and an introduction to various methods of solution of differential and integral equations. In addition, Hilbert space formalism is used to develop the foundations of quantum mechanics and Hilbert space methods are applied to optimization, variational and control problems, and to problems in approximation theory, nonlinear stability, and bifurcation. One of the most important examples of a Hilbert space is the space of the Lebesgue square integrable functions. Thus, in a study of Hilbert spaces, the Lebesgue integral cannot be avoided. In several books on Hilbert spaces, the

xv

xvi

Preface to the First Edition

reader is asked to use the Lebesgue integral pretending that it is the Riemann integral. We prefer to include a chapter on the Lebesgue integral to give the motivated reader an opportunity to understand this beautiful and powerful extension of the Riemann integral. The presentation of the Lebesgue integral is based on a method discovered independently by H.M. MacNeille and Jan ´ Mikusinski. The method eliminates the necessity of introducing the measure before the integral. This feature makes the approach more direct and less abstract. Since the main tool is the absolute convergence of numerical series, the theory is accessible for senior undergraduate students. This book is appropriate for a one-semester course in functional analysis and Hilbert space theory with applications. There are two basic prerequisites for this course: linear algebra and ordinary differential equations. It is hoped that the book will prepare students for further study of advanced functional analysis and its applications. Besides, it is intended to serve as a ready reference to the reader interested in research in various areas of mathematics, physics, and engineering sciences to which the Hilbert space methods can be applied with advantage. A wide selection of examples and exercises is included, in the hope that they will serve as a testing ground for the theory and method. Finally, a special effort is made to present a large and varied number of applications to stimulate interest in the subject. The book is divided into two parts: Part I. Theory (Chapters 1–4); Part II. Applications (Chapters 5–8). The reader should be aware that Part II is not always as rigorous as Part I. The first chapter discusses briefly the basic algebraic concepts of linear algebra and then develops the theory of normed spaces to some extent. This chapter is by no means a replacement for a course on normed spaces. Our intent was to provide the reader who has no previous experience in the theory of normed spaces with enough background for understanding of the theory of Hilbert spaces. In this chapter, we discuss normed spaces, Banach spaces, and bounded linear mappings. A section on the contraction mapping and the fixed point theorem is also included. In Chapter 2, we discuss the definition of the Lebesgue integral and prove the fundamental convergence theorems. The results are first stated and proved for real valued functions of a single variable, and then they are extended to complex valued functions of several real variables. A discussion of locally integrable functions, measure, and measurable functions is also included. In the last section, we prove some basic properties of convolution. Inner product spaces, Hilbert spaces, and orthonormal systems are discussed in Chapter 3. This is followed by discussions of strong and weak convergence, orthogonal complements and projection theorems, linear functionals, and the Riesz representation theorem. Chapter 4 is devoted to the theory of linear operators on Hilbert spaces with special emphasis on different kinds of operators and their basic properties. Bi-

Preface to the First Edition

xvii

linear functionals and quadratic forms leading to the Lax–Milgram theorem are discussed. In addition, eigenvalues and eigenvectors of linear operators are studied in some detail. These concepts play a central role in the theory of operators and their applications. The spectral theorem for self-adjoint compact operators and other related results are presented. This is followed by a brief discussion on the Fourier transforms. The last section is a short introduction to unbounded operators in a Hilbert space. Applications of the theory of Hilbert spaces to integral and differential equations are presented in Chapter 5, and emphasis is placed on basic existence theorems and the solvability of various kinds of integral equations. Ordinary differential equations, differential operators, inverse differential operators, and Green’s functions are discussed in some detail. Also included is the theory of Sturm–Liouville systems. The last section contains several examples of applications of Fourier transforms to ordinary differential equations and to integral equations. Chapter 6 provides a short introduction to distributions and their properties. The major part of this chapter is concerned with applications of Hilbert space methods to partial differential equations. Special emphasis is given to weak solutions of elliptic boundary problems, and the use of Fourier transforms for solving partial differential equations, and, in particular, for calculating Green’s functions. In Chapter 7, the mathematical foundations of quantum mechanics are built upon the theory of Hermitian operators in a Hilbert space. This chapter includes basic concepts and equations of classical mechanics, fundamental ideas and postulates of quantum mechanics, the Heisenberg uncertainty principle, the Schrödinger and the Heisenberg pictures, and the quantum theory of the linear harmonic oscillator and of the angular momentum operators. The final chapter is devoted to the Hilbert space methods for finding solutions of optimization problems, variational problems and variational inequalities, minimization problems of a quadratic functional, and optimal control problems for dynamical systems. Also included are brief treatments of approximation theory, linear and nonlinear stability problems, and bifurcation theory. This book contains almost 600 examples and exercises that are either directly associated with applications or phrased in terms of the mathematical, physical, and engineering contexts in which theory arises. The exercises truly complement the text. Answers and hints to some of them are provided at the end of the book. For students and readers wishing to learn more about the subject, important references are listed in the bibliography. In preparing this book, we have been encouraged by and have benefited from the helpful comments and criticisms of a number of graduate students and faculty members of several universities in the United States and abroad. Professors James V. Herod and Thomas D. Morley have adopted the man-

xviii

Preface to the First Edition

uscript at Georgia Institute of Technology for a graduate course on Hilbert spaces. We express our grateful thanks to them for their valuable advice and suggestions during the preparation of the book. We also wish to thank ´ Drs. R. Ger and A. Szymanski, who have carefully read parts of the manuscript and given some suggestions for improvement. It is our pleasure to acknowledge the encouragement and help of Professor P.K. Ghosh, who has provided several references and books on the subject from his personal library. We also express our grateful thanks to our friends and colleagues, including Drs. Ram N. Mohapatra, Michael D. Taylor, and Carroll A. Webber, for their interest and help during the preparation of the book. Thanks also go to Mrs. Grazyna ´ for drawing all diagrams. In spite of all the best efforts of everyone Mikusinski involved, it is doubtless that there are still typographical errors in the book. We do hope that any remaining errors are both few and obvious and will not create undue confusion. Finally, the authors wish to express their thanks to Mrs. Alice Peters, editor, and the staff of Academic Press for their help and cooperation. ´ Lokenath Debnath, Piotr Mikusinski University of Central Florida, Orlando

Normed Vector Spaces CHAPTER

1

“The organic unity of mathematics is inherent in the nature of this science, for mathematics is the foundation of all exact knowledge of natural phenomena.” David Hilbert

1.1 Introduction The basic algebraic concepts in the theory of Hilbert spaces are those of a vector space and an inner product. The inner product defines a norm, and thus every Hilbert space is a normed space. Since the norm plays a very important role in the theory, it is not possible to study Hilbert spaces without familiarity with basic concepts and properties of normed spaces. Section 1.2 provides a brief discussion of the algebraic structure of vector spaces. Section 1.3 discusses topological aspects of normed spaces. Basic properties of complete normed spaces are presented in Section 1.4. The final two sections of this chapter are devoted to linear mappings in normed spaces and the fixed point theorem. This chapter is by no means a substitute for a course in the theory of normed spaces. We limit our discussion to those concepts which are necessary for understanding of the following chapters. Some of the definitions and theorems do not require the algebraic structure of vector spaces and can be formulated in metric spaces or general topological spaces. We choose not to make these distinctions and always assume that we are dealing with a vector space or a normed vector space.

1

2

Chapter 1

Normed Vector Spaces

1.2 Vector Spaces We consider both real and complex vector spaces. The field of real numbers is denoted by R and the field of complex numbers by C. Elements of R or C are called scalars. Sometimes it is convenient to give a definition or state a theorem without specifying the field of scalars. In such a case we use F to denote either R or C. For instance, if F is used in a theorem, this means that the theorem is true for both scalar fields R and C.

Definition 1.2.1.

(Vector space) By a vector space we mean a nonempty set E with two operations: (x, y) → x + y from E × E into E called addition, (λ, x) → λx from F × E into E called multiplication by scalars,

such that the following conditions are satisfied for all x, y, z ∈ E and α, β ∈ F: (a) x + y = y + x; (b) (x + y) + z = x + (y + z); (c) For every x, y ∈ E there exists a z ∈ E such that x + z = y; (d) α(βx) = (αβ)x; (e) (α + β)x = αx + βx; (f) α(x + y) = αx + αy; (g) 1x = x. Elements of E are called vectors. If F = R, then E is called a real vector space, and if F = C, E is called a complex vector space. From (c) it follows that for every x ∈ E there exists a zx ∈ E such that x + zx = x. We will show that there exists exactly one element z ∈ E such that x + z = x for all x ∈ E. That element is denoted by 0 and called the zero vector. Let x, y ∈ E. By (c), there exists a w ∈ E such that x + w = y. If x + zx = x for some zx ∈ E, then, by (a) and (b), y + zx = (x + w) + zx = (x + zx ) + w = x + w = y. This shows that, if x + z = x for some x ∈ E, then y + z = y for any other element y ∈ E. We still need to show that such an element is unique. Indeed, if z1 and z2 are two such elements, then z1 + z2 = z1 and z1 + z2 = z2 . Thus, z 1 = z2 .

1.2 Vector Spaces

3

A similar argument shows that the vector z in (c) is unique for any pair of vectors x, y ∈ E. The unique solution z of x + z = y is denoted by y − x. According to the definition of the zero vector, we have x − x = 0. The vector 0 − x is denoted by −x. We use 0 to denote both the scalar 0 and the zero vector; this will not cause any confusion. Because of (b), the use of parentheses in expressions with more than one plus sign can be avoided. The following properties follow easily from the definition of vector spaces: If λ = 0 and λx = 0,

then x = 0.

If x = 0 and λx = 0,

then λ = 0.

0x = 0

and (−1)x = −x.

Example 1.2.2. The set {0} is a vector space. The scalar fields R and C are the simplest nontrivial vector spaces. R is a real vector space; C can be treated as a real or a complex vector space. Here are some other simple examples of vector spaces:
 RN = (x1 , . . . , xN ): x1 , . . . , xN ∈ R ,
 CN = (z1 , . . . , zN ): z1 , . . . , zN ∈ C , 
(z1 , z2 , z1 + z2 ): z1 , z2 ∈ C .




Example 1.2.3. (Function spaces) Let X be an arbitrary nonempty set and let E be a vector space. Denote by F the space of all functions from X into E. Then F becomes a vector space if the addition and multiplication by scalars are defined in the following natural way: (f + g)(x) = f (x) + g(x), (λf )(x) = λf (x). The zero vector in F is the function which assigns the zero vector of E to every element of X. Some of the most important and interesting examples of vector spaces are function spaces. We are going to encounter some of them in the first part of this book as illustrative examples and then in the second part as the natural setting for applications. Note that spaces RN and CN can be defined as function spaces: RN is the space of all real valued functions defined on {1, . . . , N} and CN is the space of all complex valued function defined on {1, . . . , N}.


4

Chapter 1

Normed Vector Spaces

A subset E1 of a vector space E is called a vector subspace (or simply a subspace if for every α, β ∈ F and x, y ∈ E1 the vector αx + βy is in E1 ). Note that a subspace of a vector space is a vector space itself. According to the definition, a vector space is a subspace of itself. If we want to exclude this case, we say a proper subspace, that is, E1 is a proper subspace of E if E1 is a subspace of E and E1 = E.

Example 1.2.4. Let  be an open subset of RN . The following are subspaces

of the space of all functions from  into C:

C () = the space of all continuous complex valued functions defined on . C k () = the space of all complex valued functions defined on  with continuous partial derivatives of order k.

C ∞ () = the space of infinitely differentiable functions defined on . P () = the space of all polynomials of N variables (considered as functions on ).




If E1 and E2 are subspaces of a vector space E and E1 ⊆ E2 , then E1 is a subspace of E2 . For instance, the space of all polynomials of N variables is a subspace of C ∞ (RN ), which in turn is a subspace of C k (RN ) or C (RN ). The intersection of subspaces of a vector space is always a subspace, but the same is not true for the union of subspaces.

Example 1.2.5. (Sequence spaces) If the set X in Example 1.2.3 is the set N of all positive integers, then the corresponding function space is a space of sequences. The addition and multiplication by scalars are then defined as (x1 , x2 , . . .) + (y1 , y2 , . . .) = (x1 + y1 , x2 + y2 , . . .), λ(x1 , x2 , . . .) = (λx1 , λx2 , . . .). The space of all sequences of complex numbers is a vector space. The space of all bounded complex sequences is a proper subspace of that space. The space of all convergent sequences of complex numbers is a proper subspace of the space of all bounded sequences.
We use the notation (xn ) or (x1 , x2 , . . .) to denote a sequence whose nth term is xn , and {xn : n ∈ N} to denote the set of all elements of the sequence. Note that {xn : n ∈ N} can be a finite set even though (xn ) is an infinite sequence. In most examples, verifying that a set is a vector space is easy or trivial. In the following example, the task is much more difficult.

Example 1.2.6. (lp -spaces) Denote by lp , for p ≥ 1, the space of all infinite

sequences (zn ) of complex numbers such that

∞ n=1

| zn |p < ∞.

5

1.2 Vector Spaces

We are going to show that lp is a vector space. Since lp is a subset of a vector space, namely the space of all sequences of complex numbers, it is enough to show that if (xn ), (yn ) ∈ lp and λ ∈ C, then (xn + yn ) ∈ lp and (λxn ) ∈ lp . To check the second property it suffices to note that ∞ 

|λxn |p = |λ|p

n=1

∞ 

|xn |p < ∞.

n=1

∞

Condition n=1 |xn + yn |p < ∞ follows immediately from Minkowski’s inequality (Hermann Minkowski (1864–1909)): 

∞ 

1/p |xn + yn |p

 ≤

n=1

1/p

∞ 

|xn |p

 +

n=1

∞ 

1/p |yn |p

.

n=1

The proof of Minkowski’s inequality is based on Hölder’s inequality (Otto Ludwig Hölder (1859–1937)). Both inequalities are proved next.


Theorem 1.2.7. (Hölder’s inequality) Let p > 1, q > 1 and p1 + 1q = 1. If (xn ) ∈

lp and (yn ) ∈ lq , then

∞ 

 |xn yn | ≤

n=1

∞ 

1/p  |xn |

p

n=1

∞ 

1/q |yn |

q

.

n=1

Proof: Without loss of generality we can assume that  ∞ n=1 |yn | = 0. First observe that

∞

n=1 |xn |

= 0 and

1 1 x1/p ≤ x + p q for 0 ≤ x ≤ 1. Let a and b be non-negative numbers such that ap ≤ bq . Then 0 ≤ ap /bq ≤ 1 and hence we have ab−q/p ≤

1 ap 1 + . p bq q

Since − pq = 1 − q, we obtain ab1−q ≤

1 ap 1 + . p bq q

Multiplying both sides by bq , we get ab ≤

ap bq + . p q

(1.1)

6

Chapter 1

Normed Vector Spaces

We have proved (1.1) assuming ap ≤ bq . A similar argument shows that (1.1) holds also if bq ≤ ap . Therefore the inequality can be used for any a, b ≥ 0. Using (1.1) with |xj | a = n ( k=1 |xk |p )1/p

|yj | b = n , ( k=1 |yk |q )1/q

and

where n ∈ N, we get

(

|yj | |xj | |xj |p |yj |q 1 1    ≤ + p 1/p ( n |y |q )1/q p nk=1 |xk |p q nk=1 |yk |q k=1 |xk | ) k=1 k

n

for any 1 ≤ j ≤ n. By adding these inequalities for j = 1, . . . , n, we obtain n

(

k=1 |xj ||yj | n p 1/p q 1/q k=1 |xk | ) ( k=1 |yk | )

n

1 1 + = 1, p q

≤

which, by letting n → ∞, gives Hölder’s inequality.




Theorem 1.2.8. (Minkowski’s inequality) Let p ≥ 1. If (xn ), (yn ) ∈ lp , then 

∞ 

1/p |xn + yn |

 ≤

p

∞ 

n=1

1/p |xn |

 +

p

n=1

∞ 

1/p |yn |

p

.

n=1

Proof: For p = 1 it is enough to use the triangle inequality for the absolute

value. If p > 1, then there exists a q such that 1/p + 1/q = 1. Then, by Hölder’s inequality, we have ∞ 

|xn + yn |p =

n=1

∞ 

|xn + yn ||xn + yn |p−1

n=1

≤

∞ 

|xn ||xn + yn |p−1 +

n=1

 ≤

∞ 

1/p  |xn |

p

n=1



+

∞  n=1

Since q(p − 1) = p,

∞ 

|yn ||xn + yn |p−1

n=1 ∞ 

1/q

|xn + yn |

q(p−1)

n=1

1/p  |yn |

p

∞  n=1

1/q |xn + yn |

q(p−1)

.

7

1.2 Vector Spaces ∞ 

 |xn + yn | ≤ p

n=1

∞ 

1/p |xn |

p

 +

n=1

∞ 

1/p  |yn |

p

n=1

∞ 

1/q |xn + yn |

p

,

n=1

which gives Minkowski’s inequality.




Example 1.2.9. (Cartesian product of vector spaces) Let E1 , . . . , En be vector spaces over the same scalar field F. Define
 E = (x1 , . . . , xn ): x1 ∈ E1 , x2 ∈ E2 , . . . , xn ∈ En with operations (x1 , . . . , xn ) + (y1 , . . . , yn ) = (x1 + y1 , . . . , xn + yn ), λ(x1 , . . . , xn ) = (λx1 , . . . , λxn ). Then E is a vector space, which is called the Cartesian product or just the product of spaces E1 , . . . , En . The product space is denoted by E = E1 × · · · × En . This idea can be generalized to an arbitrary family of vector spaces. Let J be an index set, and let Ej be a vector space for every j ∈ J. The product space E= Ej j∈J

is defined as the space of all functions x : J → j∈J Ej such that x(j) ∈ Xj for every j ∈ J. The vector space operations in E are defined as in Example 1.2.3.
Let x1 , . . . , xk be elements of a vector space E. A vector x ∈ E is called a linear combination of vectors x1 , . . . , xk if there exist scalars α1 , . . . , αk such that x = α1 x 1 + · · · + αk x k . For example, any element of RN is a linear combination of vectors e1 = (1, 0, 0, . . . , 0),

e2 = (0, 1, 0, . . . , 0),

...,

eN = (0, 0, . . . , 0, 1).

Similarly, any polynomial of degree k is a linear combination of monomials 1, x, x2 , . . . , xk . A finite collection of vectors {x1 , . . . , xk } is called linearly independent if α1 x1 + · · · + αk xk = 0 implies α1 = α2 = · · · = αk = 0. An infinite collection of vectors A is called linearly independent if every finite subcollection of A is linearly independent. A collection of vectors which is not linearly independent is called linearly dependent. We can also say that A is linearly independent if no vector x in A is a linear combination of a finite number of vectors from A different from x. Vectors e1 , . . . , eN , mentioned above, are linearly independent. Also the monomials 1, x, x2 , . . . , xk , . . . are linearly independent.

8

Chapter 1

Normed Vector Spaces

Note that linear independence may depend on the scalar field. √ For instance, the numbers 1 and i (i stands for the imaginary number i = −1) represent linearly independent vectors in the space C over the field of real numbers. On the other hand, 1 and i are not independent in C over the field of complex numbers. Let A be a subset of a vector space E. By span A we denote the set of all finite linear combinations of vectors from A, that is, span A = {α1 x1 + · · · + αk xk : x1 , . . . , xk ∈ A, α1 , . . . , αk ∈ F, k ∈ N}. It is easy to check that span A is a vector subspace of E. This subspace is called the space spanned by A. It is the smallest vector subspace of E containing A. A set of vectors B ⊂ E is called a basis of E if B is linearly independent and span B = E. If there exists a finite basis in E, then E is called a finite dimensional vector space. Otherwise we say that E is infinite dimensional. It can be proved that, for a given finite dimensional space E, the number of vectors in any basis of E is the same. If, for example, E has a basis that consists of exactly n vectors, then any other basis has exactly n vectors. In such a case n is called the dimension of E and we write dim E = n. The following are examples of sets of vectors which are bases in R3 :
 A = (1, 0, 0), (0, 1, 0), (0, 0, 1) , (1.2)
 B = (0, 1, 1), (1, 0, 1), (1, 1, 0) , (1.3)
 C = (1, 2, 3), (1, 3, 5), (3, 2, 3) . (1.4) We have dim R3 = 3, and in general dim RN = N. Spaces C (), C k (RN ), C (RN ) are infinite dimensional. Note that the dimension of the real vector space CN is 2N, while the dimension of the complex vector space CN is N. ∞

1.3 Normed Spaces In general, it does not make sense to ask what is the length or magnitude of a vector in a vector space. The concept of a norm is an abstract generalization of the length of a vector. It is defined axiomatically, that is, any real valued function satisfying certain conditions is called a norm.

Definition 1.3.1.

(Norm) A function x → x from a vector space E into R is called a norm if it satisfies the following conditions:

1.3 Normed Spaces

9

(a) x = 0 implies x = 0; (b) λx = |λ| x for every x ∈ E and λ ∈ F; (c) x + y ≤ x + y for every x, y ∈ E. Condition (c) is called the triangle inequality. Since 0 = 0 = x − x ≤ x + −x = 2 x , we have x ≥ 0 for every x ∈ E. Note that (b) implies that 0 = 0. The inequality    x − y  ≤ x − y , which follows from (b) and (c), is often used.

Example 1.3.2. The function z =



|z1 |2 + · · · + |zN |2 ,

z = (z1 , . . . , zN ) ∈ CN

defines a norm on CN . This norm is called the Euclidean norm. The following are also norms on CN : z = |z1 | + · · · + |zN |,
 z = max |z1 |, . . . , |zN | .

(1.5) (1.6)


Example 1.3.3. Let  be a closed bounded subset of R . The function f = N

maxx∈ |f (x)| defines a norm on C ().




Example 1.3.4. Let z = (zn ) ∈ lp . The function defined by z = 

p 1/p ( ∞ is a norm on lp for any p ≥ 1. Note that Minkowski’s inequality n=1 |zn | ) is in fact the triangle inequality for this norm.


Definition 1.3.5. (Normed space) A vector space with a norm is called a normed space. It is possible to define different norms on the same vector space. Therefore to define a normed space we need to specify both the vector space and the norm. We say that a normed space is a pair (E, · ), where E is a vector space and · is a norm defined on E. Some vector spaces are traditionally equipped with standard norms. For instance, when we say “the normed space RN ” we mean the Euclidean norm

2 x = x12 + · · · + xN .

10

Chapter 1

Normed Vector Spaces

Similarly, the norms defined in Examples 1.3.3 and 1.3.4 are standard. If we want to consider different norms on those spaces, we have to say something like “consider the space . . . with the norm defined by . . . .” Note that a vector subspace of a normed space is a normed space with the same norm restricted to the subspace. The absolute value is a norm on R or C. It can be used to define convergence, because the absolute value of a difference of two numbers is the distance between those numbers and convergence is about “getting closer to the limit point.” The norm plays a similar role. While x can be interpreted as the magnitude of x, x − y provides a measure of the distance between x and y.

Definition 1.3.6.

(Convergence in a normed space) Let (E, · ) be a normed space. We say that a sequence (xn ) of elements of E converges to some x ∈ E, if for every ε > 0 there exists a number M such that for every n ≥ M we have xn − x < ε . In such a case we write limn→∞ xn = x or simply xn → x . This definition becomes simpler if convergence of real numbers is used: xn → x in E means xn − x → 0 in R. The convergence in a normed space has the basic properties of convergence in R: A convergent sequence has a unique limit. If xn → x and λn → λ (λn , λ are scalars), If xn → x and yn → y,

then λn xn → λx.

then xn + yn → x + y.

These properties can be proved the same way as in the case of convergence in R. A norm on a vector space E induces a convergence in E. In other words, if we have a normed space E, then we automatically have a convergence defined in E. In practice, we often face a different situation: we have a vector space E with a given convergence and we want to know if we can find a norm on E, which would define that convergence. It is not always possible. The following two examples illustrate the problem. In the first one, the convergence can be described by a norm. In the second one, we prove that the given convergence cannot be defined by a norm.

Example 1.3.7. (Uniform convergence) Consider the space C () of all continuous functions defined on a closed bounded set  ⊂ RN . We say that the sequence f1 , f2 , . . . ∈ C () converges uniformly to f if for every ε > 0 there exists a constant n0 such that |f (x) − fn (x)| < ε for all x ∈  and all n ≥ n0 . It is easy to see that the norm in Example 1.3.3 defines uniform convergence, that is, the sequence (fn ) converges uniformly to f if and only if

1.3 Normed Spaces

11

Figure 1.1 A typical function gn in Example 1.3.8. fn − f = maxx∈ |fn (x) − f (x)| → 0 as n → ∞. For this reason this norm is called the uniform convergence norm.


Example 1.3.8. (Pointwise convergence) Let f , f1 , f2 , . . . ∈ C ([0, 1]), the space of all continuous functions on the interval [0, 1]. We say that the sequence (fn ) is pointwise convergent to f if |fn (t) − f (t)| → 0, as n → ∞, for every t ∈ [0, 1]. We will show that there is no norm on C ([0, 1]), which defines pointwise convergence. Suppose, on the contrary, that · is a norm on C ([0, 1]) such that fn − f → 0 if and only if fn (t) → f (t) for every t ∈ [0, 1].

(1.7)

Consider the sequence of functions g1 , g2 , . . . defined by  n 2 t gn (t) = 2 − 2n t  0

if 0 ≤ t ≤ 2−n , if 2−n ≤ t ≤ 21−n , otherwise

(see Figure 1.1). Since gn = 0, we have gn = 0 for all n ∈ N. If we define fn = gn / gn , then fn = 1 for all n ∈ N, and thus the sequence (fn ) is not convergent to 0 (the zero function) with respect to the norm · . On the other hand, it is easy to see that fn (t) → 0 for every t ∈ [0, 1]. This contradiction shows that a norm satisfying (1.7) cannot exist.


Definition 1.3.9. (Equivalence of norms) Two norms defined on the same vector space are called equivalent if they define the same convergence. More precisely, norms · 1 and · 2 on E are equivalent if for any sequence (xn ) in E and any x ∈ E, xn − x 1 → 0

if and only if xn − x 2 → 0.

12

Chapter 1

Normed Vector Spaces

Example 1.3.10. The following norms on R2 are equivalent:     (x, y) = x2 + y2 , (x, y) = |x| + |y|, 1 2  
 (x, y) = max |x|, |y| . 3




We will prove that any two norms on a finite dimensional vector space are equivalent. The proof requires some preparations. First we prove a useful criterion for equivalence of norms. The condition in the following theorem is often used as a definition of equivalence of norms.

Theorem 1.3.11. (Equivalence of norms) Let · 1 and · 2 be norms on a vector space E. Then · 1 and · 2 are equivalent if and only if there exist positive numbers α and β such that α x 1 ≤ x 2 ≤ β x 1

for all x ∈ E.

Proof: Clearly, the above condition implies equivalence of norms · 1 and · 2 . Now assume that the norms are equivalent, that is, xn − x 1 → 0 if and only if xn − x 2 → 0. Suppose there is no α > 0 such that α x 1 ≤ x 2 for every x ∈ E. Then, for each n ∈ N, there exists xn ∈ E such that 1 xn 1 > xn 2 . n Define 1 xn yn = √ . n xn 2

√ Then yn 2 = √1n → 0. On the other hand, yn 1 ≥ n yn 2 ≥ n. This contradiction shows that a number α with the required property exists. Existence of the number β can be proved in a similar way.
The following technical lemma establishes a property crucial for the proof of equivalence of norms on finite dimensional spaces.

Lemma 1.3.12. If x1 , . . . , xn are linearly independent elements of a normed space E, then there exists a constant c > 0 such that   (1.8) α1 x1 + · · · + αn xn ≥ c |α1 | + · · · + |αn | for all α1 , . . . , αn ∈ R.

Proof: Since (1.8) holds for any c if |α1 | + · · · + |αn | = 0, we can assume that |α1 | + · · · + |αn | > 0. Then (1.8) is equivalent to β1 x1 + · · · + βn xn ≥ c

whenever |β1 | + · · · + |βn | = 1.

(1.9)

1.3 Normed Spaces

13

The function f : Rn → R defined by f (β1 , . . . , βn ) = β1 x1 + · · · + βn xn is continuous. Since the set B = {(β1 , . . . , βn ) ∈ Rn : |β1 | + · · · + |βn | = 1} is closed and bounded, f attains a minimum on B. Note that the minimum cannot be 0, because then we would have β1 x1 + · · · + βn xn = 0 for some |β1 | + · · · + |βn | = 1, contradicting linear independence of x1 , . . . , xn . Therefore, c=

min

(β1 ,...,βn )∈B

f (β1 , . . . , βn ) =

min

(β1 ,...,βn )∈B

β1 x1 + · · · + βn xn > 0,

proving (1.9).




Theorem 1.3.13. If E is finite dimensional, then any two norms on E are equivalent.

Proof: Let E be a finite dimensional vector space, and let {e1 , . . . , en } be a basis in E. We will prove that an arbitrary norm · in E is equivalent to the norm · 0 defined by α1 e1 + · · · + αn en 0 = |α1 | + · · · + |αn |. We use Theorem 1.3.11. First note that α1 e1 + · · · + αn en ≤ |α1 | e1 + · · · + |αn | en
  ≤ max e1 , . . . , en |α1 | + · · · + |αn | . Therefore, for any x ∈ E, x ≤ β x 0 , where β = max{ e1 , . . . , en }. Existence of a constant α > 0 such that α x 0 ≤ x , for any x ∈ E, follows immediately from Lemma 1.3.12.




A normed space (E, · ) becomes a metric space if we define a metric by d(x, y) = x − y . The convergence defined by the norm and the convergence defined by this metric are the same. The metric in E defines a topology in E and thus all topological notions, that is, notions which can be defined in terms of open sets. It is not necessary to define the metric first and then the topology using that metric. In the remainder of this section, we define the basic topological notions in terms of the norm. We also prove some topological properties of normed spaces.

14

Chapter 1

Normed Vector Spaces

Figure 1.2 Examples of unit balls in R2 (Example 1.3.15).

Definition 1.3.14.

(Open ball, closed ball, sphere) Let x be an element of a normed space E, and let r be a positive number. We use the following notation:
 B(x, r) = y ∈ E: y − x < r (open ball);
 B(x, r) = y ∈ E: y − x ≤ r (closed ball);
 S(x, r) = y ∈ E: y − x = r (sphere).

In each case, x is called the center and r the radius.

Example 1.3.15. Figure 1.2 shows examples of balls in R2 with respect to the norms

    (x, y) = x2 + y2 , (x, y) = |x| + |y|, 1 2  
 (x, y) = max |x|, |y| . 3




Example 1.3.16. Let E = C ([−π, π ]), and let f = maxt∈[−π,π ] |f (t)|. Figure 1.3 shows B({sin t}, 1). The figure should be understood as follows: B({sin t}, 1) is the set of all continuous functions on [−π, π ] whose graphs are in the shaded area. It is not the set of all points of the shaded area.


Definition 1.3.17.

(Open and closed sets) A subset S of a normed space E is called open if for every x ∈ S there exists ε > 0 such that B(x, ε) ⊆ S. A subset S is called closed if its complement is open, that is, if E\S is open. It is important to realize that equivalent norms define the same open sets, even though the balls are different. The same is true for closed, dense, compact sets, and other notions that can be defined in terms of open sets.

1.3 Normed Spaces

15

Figure 1.3 The ball B({sin t}, 1) in C ([−π, π ]).

Example 1.3.18. Open balls are open sets. Closed balls and spheres are closed sets.




Example 1.3.19. Let  be a closed bounded set in RN . Consider the space

C () with the norm f = maxx∈ |f (x)|. Let f ∈ C () and f (x) > 0 for all x ∈ . The following sets are open in C (): 
g ∈ C (): g(x) < f (x) for all x ∈  , 
g ∈ C (): g(x) > f (x) for all x ∈  ,   
g ∈ C (): g(x) < f (x) for all x ∈  ,   
g ∈ C (): g(x) > f (x) for all x ∈  . The following sets are closed in C (): 
g ∈ C (): g(x) ≤ f (x) for all x ∈  , 
g ∈ C (): g(x) ≥ f (x) for all x ∈  ,   
g ∈ C (): g(x) ≤ f (x) for all x ∈  ,   
g ∈ C (): g(x) ≥ f (x) for all x ∈  , 
g ∈ C (): g(x0 ) = λ (x0 is a fixed point in  and λ ∈ C).


Theorem 1.3.20. (a) The union of any collection of open sets is open. (b) The intersection of a finite number of open sets is open. (c) The union of a finite number of closed sets is closed.

16

Chapter 1

Normed Vector Spaces

(d) The intersection of any collection of closed sets is closed. (e) The empty set and the whole space are both open and closed. The proofs are left as exercises.

Theorem 1.3.21. A subset S of a normed space E is closed if and only if every sequence of elements of S convergent in E has its limit in S, that is, x1 , x2 , . . . ∈ S and xn → x

implies x ∈ S.

Proof: Suppose S is a closed subset of E, x1 , x2 , . . . ∈ S, xn → x, and x ∈/ S. Since S is closed, E\S is open. Thus, there exists an ε > 0 such that B(x, ε) ⊆ E\S. On the other hand, since x − xn → 0, we have x − xn < ε for all sufficiently large n ∈ N. This contradiction shows that x ∈ S. Suppose now that whenever x1 , x2 , . . . ∈ S and xn → x, then x ∈ S. If S is not closed, then E\S is not open, and thus there exists x ∈ E\S such that every ball B(x, ε) contains elements of S. Thus, we can find x1 , x2 , . . . ∈ S such that xn ∈ B(x, 1/n). But then xn → x and, according to our assumption, x ∈ S. This contradicts the assumption x ∈ E\S. Therefore S must be a closed set.
Note how the above theorem is useful in proving that the sets mentioned in Example 1.3.19 are closed.

Definition 1.3.22.

(Closure) Let S be a subset of a normed space E. By the closure of S, denoted by cl S, we mean the intersection of all closed sets containing S.

In view of Theorem 1.3.20(d), the closure of a set is always a closed set. It is the smallest closed set which contains S. The following theorem gives a sequential description of the closure.

Theorem 1.3.23. Let S be a subset of a normed space E. The closure of S is the set of limits of all convergent sequences of elements of S, that is, cl S = {x ∈ E: there exist x1 , x2 , . . . ∈ S such that xn → x}. The proof is left as an exercise.

Example 1.3.24. The Weierstrass theorem says that every continuous function on an interval [a, b] can be uniformly approximated by polynomials. This can also be expressed as follows: The closure of the set of all polynomials on [a, b] is the whole space C ([a, b]).


1.3 Normed Spaces

17

Definition 1.3.25.

(Dense subset) A subset S of a normed space E is called dense in E if cl S = E.

Example 1.3.26. The set of all polynomials on [a, b] is dense in C ([a, b]). The set of all sequences of complex numbers, which have only a finite number of nonzero terms is dense in lp for all p ≥ 1.


Theorem 1.3.27. Let S be a subset of a normed space E. The following conditions are equivalent: (a) S is dense in E, (b) For every x ∈ E there exist x1 , x2 , . . . ∈ S such that xn → x, (c) Every nonempty open subset of E contains an element of S. The proof is left as an exercise.

Definition 1.3.28.

(Compact set) A subset S of a normed space E is called compact if every sequence (xn ) in S contains a convergent subsequence whose limit belongs to S.

Example 1.3.29. In RN or CN , a set is compact if and only if it is bounded and closed.




This simple characterization of compact sets is valid for any finite dimensional normed spaces. In infinite dimensional normed spaces the situation is more complicated.

Definition 1.3.30.

(Bounded subset) A subset S of a normed space E is called bounded if S ⊆ B(0, r) for some r > 0.

It is easy to show that S is bounded if and only if λn xn → 0 for any xn ∈ S and any scalars λn → 0.

Theorem 1.3.31. Compact sets are closed and bounded. Proof: Let S be a compact subset of a normed space E. Suppose x1 , x2 , . . . ∈ S and xn → x. Then (xn ) contains a subsequence (xpn ), which converges to some y ∈ S. On the other hand, we have xpn → x. Thus, x = y and x ∈ S. This shows that S is closed.

18

Chapter 1

Normed Vector Spaces

Suppose now that S is not bounded. Then there exists a sequence x1 , x2 , . . . ∈ S such that xn ≥ n for all n ∈ N. Clearly, (xn ) does not contain a convergent subsequence, and hence S is not compact.
The next example shows that closed and bounded sets in infinite dimensional spaces need not be compact.

Example 1.3.32. Consider the space C ([0, 1]). The closed unit ball B(0, 1) is a closed and bounded set, but it is not compact. To see this consider the sequence of functions defined by xn (t) = t n . Then xn ∈ B(0, 1) for all n ∈ N. Clearly, no subsequence of (xn ) converges in C ([0, 1]).
It turns out that finite dimensional normed spaces can be characterized by the property that the unit ball is compact. The proof of this fact is based on a theorem that is usually referred to as Riesz’s lemma.

Theorem 1.3.33. (Riesz’s lemma) Let X be a closed proper subspace of a normed space E. For every ε ∈ (0, 1) there exists an xε ∈ E such that xε = 1 and xε − x ≥ ε for all x ∈ X.

Proof: Let z ∈ E\X and let d = infx∈X z − x . Since X is closed, d > 0. If

ε ∈ (0, 1), then

d ≤ z − x0 ≤

d ε

for some x0 ∈ X. We will show that xε =

z − x0 z − x0

has the desired property. Indeed, for any x ∈ X we have    z − x0   − x xε − x =   z − x0   1 z − x0 − z − x0 x = z − x0 ε 1 d ≥ d = ε, ≥ z − x0 d where the first inequality follows from the definition of d since x0 + z − x0 x ∈ X.


Theorem 1.3.34. A normed space E is finite dimensional if and only if the closed unit ball in E is compact. Proof: Let E be finite dimensional and let {e1 , . . . , eN } be a basis in E. By Theorem 1.3.13, we can assume, without loss of generality, that the norm

1.4 Banach Spaces

19

on E is defined by α1 e1 + · · · + αN eN = |α1 | + · · · + |αN |. It is a simple exercise to show that the unit ball with respect to this norm is compact. Now assume that E is infinite dimensional. Using Riesz’s lemma, we construct a sequence xn ∈ E such that xn = 1 for all n ∈ N and xm − xn ≥ 12 for all m, n ∈ N, m = n. Let x1 ∈ E be any element such that x1 = 1. Since span{x1 } is a closed and proper subspace of E, there exists an x2 ∈ E such that x2 = 1 and x1 − x2 ≥ 12 . Now suppose we have already constructed x1 , . . . , xn with the desired properties. Since span{x1 , . . . , xn } is a closed and proper subspace of E, there exists an xn+1 ∈ E such that xn+1 = 1 and xk − xn+1 ≥ 12 for all k = 1, 2, . . . , n. Suppose (xn ) has a convergent subsequence (xpn ) such that xpn → x, for some x ∈ E. Then 1 ≤ xpn − xpn+1 ≤ xpn − x + xpn+1 − x 2 for all n ∈ N, which contradicts convergence of (xpn ).




1.4 Banach Spaces Every Cauchy sequence of numbers converges. Every absolutely convergent series of numbers converges. These are very important properties of real and complex numbers. Many crucial arguments concerning numbers rely on them. One expects that similar properties of a normed space would be of great importance. This is true. However, not all normed spaces have these properties. Those that do are called Banach spaces after the name of Stafan Banach (1892– 1945).

Definition 1.4.1.

(Cauchy sequence) A sequence of vectors (xn ) in a normed space is called a Cauchy sequence if for every ε > 0 there exists a number M such that xm − xn < ε for all m, n > M . Cauchy sequences are named after Augustin Louis Cauchy (1789–1857).

Theorem 1.4.2. The following conditions are equivalent: (a) (xn ) is a Cauchy sequence; (b) xpn − xqn → 0 as n → ∞, for every pair of increasing sequences of positive integers (pn ) and (qn );

20

Chapter 1

Normed Vector Spaces

(c) xpn+1 − xpn → 0 as n → ∞, for every increasing sequence of positive integers (pn ). The proof is left as an exercise. Observe that every convergent sequence is a Cauchy sequence. In fact, if xn − x → 0, then xpn − xqn ≤ xpn − x + xqn − x → 0, for every pair of increasing sequences of indices (pn ) and (qn ). The converse is not true, in general.

Example 1.4.3. Let P ([0, 1]) be the space of polynomials on [0, 1] with the norm of uniform convergence P = max[0,1] |P(x)|. Define Pn (x) = 1 + x +

x2 xn + ··· + 2! n!

for n = 1, 2, . . . . Then (Pn ) is a Cauchy sequence, but it does not converge in P ([0, 1]) because its limit is not a polynomial.


Lemma 1.4.4. If (xn ) is a Cauchy sequence in a normed space, then the sequence of norms ( xn ) converges.

Proof: Since | x − y | ≤ x −y , we have | xm − xn | ≤ xm −xn → 0 as m, n → ∞. This shows that the sequence of norms is a Cauchy sequence of real numbers, hence it is convergent.
Note that this lemma implies that every Cauchy sequence is bounded, that is, if (xn ) is a Cauchy sequence, then there is a number M such that xn ≤ M for all n.

Definition 1.4.5.

(Banach space) A normed space E is called complete if every Cauchy sequence in E converges to an element of E. A complete normed space is called a Banach space.

Example 1.4.6. We will show that the space l2 is complete. Let (an ) be a Cauchy sequence in l2 . If an = (αn,1 , αn,2 , . . .), then given any ε > 0, there exists a number n0 such that ∞  k=1

|αm,k − αn,k |2 < ε2

(1.10)

1.4 Banach Spaces

21

for all m, n ≥ n0 . Note that this implies that for every fixed k ∈ N and for every ε > 0 there exists a number n0 such that |αm,k − αn,k | < ε for all m, n ≥ n0 . But this means that, for every k, the sequence (αn,k ) is a Cauchy sequence in C and thus convergent. Denote αk = lim αn,k ,

k = 1, 2, . . .

n→∞

and

a = (αn ).

We are going to prove that a is an element of l2 and that the sequence (an ) converges to a. Indeed, from (1.10), by letting m → ∞, we obtain ∞ 

|αk − αn,k |2 ≤ ε2

(1.11)

k=1

for every n ≥ n0 . Since ∞ 

|αn0 ,k |2 < ∞,

k=1

by Minkowski’s inequality, we have   ∞ ∞    2 2  |αk | =  |αk | − |αn0 ,k | + |αn0 ,k | k=1

k=1

  ∞ ∞    2   ≤ |αn ,k |2 |αk | − |αn ,k | +  0

0

k=1

k=1

  ∞ ∞   2 ≤ |αk − αn ,k | +  |αn ,k |2 < ∞. 0

k=1

0

k=1

This proves that the sequence a = (αn ) is an element of l2 . Moreover, since ε is arbitrarily small, (1.11) implies  ∞   2 lim a − an = lim  |αk − αn,k | = 0, n→∞

n→∞

k=1

which means that the sequence (an ) is convergent to a in l2 .




Example 1.4.7. Another important example of a Banach space is the space C ([a, b]) of continuous (real or complex valued) functions on an interval [a, b]. Recall that the norm on C ([a, b]) is defined by f = max[a,b] |f (x)|.

22

Chapter 1

Normed Vector Spaces

Let (fn ) be a Cauchy sequence in C ([a, b]). For an arbitrary ε > 0 there exists an n0 ∈ N such that fn − fm < ε and thus also   fn (x) − fm (x) < ε

for all n, m ≥ n0 ,

for all n, m ≥ n0 and all x ∈ [a, b].

(1.12)

Note that this implies that (fn (x)) is a Cauchy sequence for every x ∈ [a, b]. Completeness of R (or C) allows us to define f (x) = lim fn (x), n→∞

x ∈ [a, b].

Now, by letting m → ∞ in (1.12), we get   fn (x) − f (x) ≤ ε for all n ≥ n0 and all x ∈ [a, b].

(1.13)

Let x0 ∈ [a, b]. Since fn0 is continuous on [a, b], there exists a δ > 0 such that |fn0 (x0 ) − fn0 (y)| < ε for every y ∈ [a, b] such that |x0 − y| < δ. Then         f (x0 ) − f (y) ≤ f (x0 ) − fn (x0 ) + fn (x0 ) − fn (y) + fn (y) − f (y) 0 0 0 0 < ε + ε + ε = 3ε, whenever |x0 − y| < δ. Since x0 and ε are arbitrary, the above proves continuity of f . Finally, since (1.13) implies fn − f ≤ ε

for all n ≥ n0 ,

the sequence (fn ) converges to f uniformly.




Definition  1.4.8.

(Convergent and absolutely convergent series) A series ∞ x in a normed space E is called convergent if the sequence of partial n=1 n x1 +x2 +· · ·+xn −x → sums converges in E, that is, there exists x ∈ E such that ∞ 0 as n → ∞. In that case we write ∞ x = x . If n n=1 n=1 xn < ∞, then the series is called absolutely convergent. In general, an absolutely convergent series need not converge, as can be seen from Example 1.4.3.

Theorem 1.4.9. A normed space is complete if and only if every absolutely convergent series converges.

1.4 Banach Spaces

Proof: Let E be a Banach space. Suppose xn ∈ E and

∞

n=1 xn

23 < ∞. De-

fine sn = x1 + · · · + xn ,

n = 1, 2, . . .

We will show that (sn ) is a Cauchy sequence. Let ε > 0 and let k be a positive integer such that ∞ 

xn < ε.

n=k+1

Then, for every m > n > k, we have sm − sn = xn+1 + · · · + xm ≤

∞ 

xr < ε.

r=n+1

in E. Since E is comThis proves that the sequence (sn ) is a Cauchy sequence plete, (sn ) converges in E, which means that the series ∞ n=1 xn converges. Assume now, that E is a normed space in which every absolutely convergent series converges. We need to prove that E is complete. Let (xn ) be a Cauchy sequence in E. Then, for every k ∈ N, there exists a pk ∈ N such that xm − xn < 2−k

for all m, n ≥ pk .

Without loss of generality, we  can assume that the sequence (pn ) is strictly increasing. Since the series ∞ k=1 (xpk+1 − xpk ) is absolutely convergent, it is convergent, and thus the sequence xpk = xp1 + (xp2 − xp1 ) + · · · + (xpk − xpk−1 ) converges to an element x ∈ E. Consequently, xn − x ≤ xn − xpn + xpn − x → 0, because (xn ) is a Cauchy sequence.




Theorem 1.4.10. A closed vector subspace of a Banach space is a Banach space itself.

Proof: Let (E, · ) be a Banach space and let F be a closed vector subspace of E. If (xn ) is a Cauchy sequence in F, then it is a Cauchy sequence in E, and therefore there exists an x ∈ E such that xn → x. Since F is a closed subset of E, we have x ∈ F. Thus every Cauchy sequence in F converges to an element of F.
Some spaces arising naturally in applications are not complete. It turns out that it is always possible to enlarge such a space to a complete space. The main ideas of the construction are described next.

24

Chapter 1

Normed Vector Spaces

˜ · 1 ) is called a comLet (E, · ) be a normed space. A normed space (E, pletion of (E, · ) if (a) There exists a one-to-one mapping  : E → E˜ such that (αx + βy) = α(x) + β(y)

for all x, y ∈ E and α, β ∈ F,

(b) x = (x) 1 for every x ∈ E, ˜ (c) (E) is dense in E, (d) E˜ is complete. ˜ satisfying these conditions, can be formally defined as the space A space E, of equivalence classes of Cauchy sequences of elements of E. Two Cauchy sequences (xn ) and (yn ) of elements of E are called equivalent if limn→∞ xn − yn = 0. The set of all Cauchy sequences equivalent to a given Cauchy sequence (xn ) is denoted by [(xn )] and called the equivalence class of (xn ). The set of ˜ all equivalence classes of Cauchy sequences of elements of E, denoted by E, becomes a vector space when the addition and multiplication by scalars are defined as follows:           (xn ) + (yn ) = (xn + yn ) and λ (xn ) = (λxn ) . The norm on E˜ is defined by   [xn ] = lim xn . 1 n→∞

By Lemma 1.4.4, limn→∞ xn exists for every Cauchy sequence (xn ). It easy to show that if [(xn )] and [(yn )] are equivalent, then limn→∞ xn = limn→∞ yn . Now define  : E → E˜ by (x) = [(x, x, . . .)]. Clearly,  satisfies (a) and (b). To show that (E) is dense in E˜ note that every element [(xn )] of E˜ is the limit of the sequence ((xn )). ˜ Now we will prove that E˜ is complete. Let (Xn ) be a Cauchy sequence in E. ˜ for every n ∈ N there exists xn ∈ E such that Since (E) is dense in E,   (xn ) − Xn  < 1 . 1 n From the inequalities   xn − xm = (xn ) − (xm )1     ≤ (xn ) − Xn 1 + Xn − Xm 1 + Xm − (xm )1 ≤ Xn − Xm 1 +

1 1 + , n m

1.5 Linear Mappings

25

we see that (xn ) is a Cauchy sequence in E. Define X = [(xn )]. It remains to show that lim Xn − X 1 = 0.

n→∞

Indeed, we have       1 Xn − X 1 ≤ Xn − (xn )1 + (xn ) − X 1 < (xn ) − X 1 + → 0, n because limn→∞ (xn ) − X 1 = 0.

1.5 Linear Mappings First, we introduce some notation that will be used in the remainder of the book. Let E1 and E2 be vector spaces, and let L be a mapping from E1 into E2 . If y = L(x), then y is called the image of x. If A is a subset of E1 , then L(A) denotes the image of the set A, that is, L(A) is the set of all elements in E2 , which are images of elements of A. If B is a subset of E2 , then L−1 (B) denotes the inverse image of B, that is, L−1 (B) is the set of all elements in E1 whose images are elements of B. In symbols:

 L(A) = L(x): x ∈ A , L−1 (B) = x ∈ E1 : L(x) ∈ B . Note that the use of notation L−1 does not imply that L is invertible. We often consider mappings, which are defined on proper subsets of vector spaces. Then it is important to specify the domain of L, denoted by D (L). The set L(D (L)) is called the range of L and denoted by R(L), that is,
 R(L) = y ∈ E2 : L(x) = y for some x ∈ D(L) . By the null space of L, denoted by N (L), we mean the set of all elements x ∈ D(L) such that L(x) = 0. Finally, by the graph of L, denoted by G (L), we mean the subset of E1 × E2 defined as follows:
 G (L) = (x, y): x ∈ D(L) and y = L(x) .

Definition 1.5.1. (Linear mapping) A mapping L : E1 → E2 is called a linear mapping if L(αx + βy) = αL(x) + βL(y) for all x, y ∈ E1 and all scalars α, β . In the context of mappings between vector spaces we often write Lx instead of L(x).

26

Chapter 1

Normed Vector Spaces

In Definition 1.5.1 we are assuming that the domain of L is a vector space. If the domain of L is a subset E1 which is not a vector space, then we have to be more careful when defining linearity of L. We can say that a mapping L from a subset S = D (L) of a vector space E1 into a vector space E2 is linear if L(αx + βy) = αLx + βLy for all x, y ∈ S and all scalars α, β such that αx + βy ∈ S. On the other hand, such a mapping L has a unique extension to a linear mapping from the vector space span D (L) into E2 . Indeed, if y ∈ span D (L), then y = λ1 x1 + · · · + λn xn , for some x1 , . . . , xn ∈ D (L) and some scalars λ1 , . . . , λn , and we can define Ly = λ1 Lx1 + · · · + λn Lxn . For this reason, we can always assume, without loss of generality, that the domain of a linear mapping is a vector space. Under this assumption, it can be easily shown that R(L), N (L), and G (L) are vector spaces. In the remaining part of this section we will assume that both spaces E1 and E2 are normed spaces. We use the same symbol · to denote the norm on E1 as well as on E2 . This will not lead to any misunderstanding.

Definition 1.5.2.

(Continuous mapping) Let E1 and E2 be normed spaces. A mapping F from E1 into E2 is called continuous at x0 ∈ E1 if, for any sequence (xn ) of elements of E1 convergent to x0 , the sequence (F(xn )) converges to F(x0 ), that is, F is continuous at x0 if xn − x0 → 0 implies F(xn ) − F(x0 ) → 0. If F is continuous at every x ∈ E1 , then we simply say that F is continuous. Several examples of continuous mappings are discussed in Chapter 4 (Section 4.2). Here we only make the following simple but useful observation.

Example 1.5.3. The norm on a normed space E is a continuous mapping from E into R. Indeed, if xn − x → 0, then | xn − x | ≤ xn − x → 0.




Continuity can be described in many different ways. The conditions in the following theorem characterize continuity in terms of open and closed sets. The proof of the theorem is left as an exercise.

Theorem 1.5.4. Let F : E1 → E2 . The following conditions are equivalent: (a) F is continuous; (b) The inverse image F −1 (U) of any open subset U of E2 is open in E1 ; (c) The inverse image F −1 (S) of any closed subset S of E2 is closed in E1 . From now on we are going to limit our discussion to linear mappings.

Theorem 1.5.5. If a linear mapping L : E1 → E2 is continuous at some x0 ∈ E1 , then it is continuous.

27

1.5 Linear Mappings

Proof: Assume L is continuous at x0 ∈ E1 . Let x be an arbitrary element

of E1 and let (xn ) be a sequence convergent to x. Then the sequence (xn − x + x0 ) converges to x0 and thus we have Lxn − Lx = L(xn − x + x0 ) − Lx0 → 0, which completes the proof.


Definition 1.5.6. (Bounded linear mapping) A linear mapping L : E1 → E2 is called bounded if there exists a number α > 0 such that Lx ≤ α x for all x ∈ E1 . Note that the condition in this definition is equivalent to saying that L is bounded by α on the unit sphere in E1 . More precisely, Lx ≤ α for all x ∈ E1 such that x = 1.

Theorem 1.5.7. A linear mapping is continuous if and only if it is bounded. Proof: If L is bounded and xn → 0, then Lxn ≤ α xn → 0. Thus, L is continuous at 0 and hence, by Theorem 1.5.5, L is continuous. If L is not bounded, then for every n ∈ N there exists an xn ∈ E1 such that Lxn > n xn . Define yn =

xn , n xn

n = 1, 2, . . . .

Then yn → 0. Since Lyn > 1 for all n ∈ N, L is not continuous.




Since every linear mapping between finite dimensional spaces is bounded, it is continuous. The above theorem implies also that, for linear mappings, continuity and uniform continuity are equivalent. The space of all linear mappings from a vector space E1 into a vector space E2 becomes a vector space if the addition and multiplication by scalars are defined as follows: (L1 + L2 )x = L1 x + L2 x

and

(λL)x = λ(Lx).

If E1 and E2 are normed spaces, then the set of all bounded linear mappings from E1 into E2 , denoted by B (E1 , E2 ), is a vector subspace of the space defined above.

Theorem 1.5.8. If E1 and E2 are normed spaces, then B(E1 , E2 ) is a normed space with the norm defined by L = sup Lx . x =1

(1.14)

28

Chapter 1

Normed Vector Spaces

Proof: We will only show that the norm (1.14) satisfies the triangle inequality. If L1 , L2 ∈ B (E1 , E2 ) and x ∈ E1 is such that x = 1, then L1 x + L2 x ≤ L1 x + L2 x . This implies L1 x + L2 x ≤ sup L1 x + sup L2 x = L1 + L2 , x =1

x =1

and hence L1 + L2 = sup L1 x + L2 x ≤ L1 + L2 . x =1




It follows from (1.14) that Lx ≤ L x for all x ∈ E1 . In fact, L is the least number α such that Lx ≤ α x for all x ∈ E1 . The norm defined in (1.14) is the standard norm on B (E1 , E2 ) and this is the norm we mean when we say “the normed space B (E1 , E2 ).” Convergence with respect to this norm is called uniform convergence.

Theorem 1.5.9. If E1 is a normed space and E2 is a Banach space, then B(E1 , E2 ) is a Banach space. Proof: We only need to show that B(E1 , E2 ) is complete. Let (Ln ) be a Cauchy sequence in B (E1 , E2 ) and let x be an arbitrary element of E1 . Then Lm x − Ln x ≤ Lm − Ln x → 0 as m, n → ∞, which shows that (Ln x) is a Cauchy sequence in E2 . By completeness of E2 , there is a unique element y ∈ E2 such that Ln x → y. Since x is an arbitrary element of E1 , this defines a mapping L from E1 into E2 : Lx = lim Ln x. n→∞

We will show that L ∈ B (E1 , E2 ) and Ln − L → 0. Clearly, L is a linear mapping. Since Cauchy sequences are bounded, there exists a constant α such that Ln ≤ α for all n ∈ N. Consequently,     Lx =  lim Ln x = lim Ln x ≤ α x . n→∞

n→∞

Therefore L is bounded and thus L ∈ B (E1 , E2 ). It remains to prove that Ln − L → 0. Let ε > 0, and let k be such that Lm − Ln < ε for every m, n ≥ k. If x = 1 and m, n ≥ k, then Lm x − Ln x ≤ Lm − Ln < ε.

1.5 Linear Mappings

29

By letting n → ∞, (m remains fixed), we obtain Lm x − Lx ≤ ε for every m ≥ k and every x ∈ E1 with x = 1. This means that Lm − L ≤ ε for all m > k, which completes the proof.


Theorem 1.5.10. If L is a continuous linear mapping from a subspace of a normed space E1 into a Banach space E2 , then L has a unique extension to a continuous linear mapping defined on the closure of the domain D (L). In particular, if D (L) is dense in E1 , then L has a unique extension to a continuous linear mapping defined on the whole space E1 .

Proof: If x ∈ cl D(L), then there exists a sequence (xn ) in D(L) convergent to x. Since (xn ) is a Cauchy sequence,   Lxm − Lxn = L(xm − xn ) ≤ L xm − xn → 0,

as m, n → ∞.

Thus, (Lxn ) is a Cauchy sequence in E2 . Since E2 is complete, there is a z ∈ E2 such that Lxn → z. We want to define the value of the extension L˜ at x as L˜ x = z, that is, L˜ x = lim Lxn , n→∞

xn ∈ D (L) and xn → x.

This definition will be correct only if we can show that the limit z is the same for all sequences in D (L) convergent to x. Indeed, if yn ∈ D (L) and yn → x, then Lyn = Lyn − Lxn + Lxn = L(yn − xn ) + Lxn → z, because yn − xn → 0, and hence also L(yn − xn ) → 0. Clearly, L˜ is a linear mapping and L˜ x = Lx whenever x ∈ D (L). It remains to show that L˜ is continuous. Let x ∈ cl D (L), x = 1. There exist x1 , x2 , . . . ∈ D (L) such that xn → x. Then xn → x = 1 and   L˜ x = lim Lxn ≤ L . n→∞

Thus, L˜ is bounded, hence continuous, and L˜ = L .




Theorem 1.5.11. If L : E1 → E2 is a continuous linear mapping, then the null space N (L) is a closed subspace of E. Moreover, if the domain D (L) is closed, then the graph G (L) is a closed subspace of E1 × E2 . The proof is left as an exercise. Spaces B (E, F) of bounded linear mappings from a normed space E into the scalar field F are of special interest. Elements of B (E, F) are called functionals.

30

Chapter 1

Normed Vector Spaces

The space B (E, F) is usually denoted by E and called the dual space of E. Theorems proved in this section apply to dual spaces of normed spaces. Note that since the scalar field is a complete space, the dual space of a normed space is always a Banach space. The last theorem we prove in this section is the Banach–Steinhaus theorem, also known as the uniform boundedness principle (Hugo Dyonizy Steinhaus (1887–1972)). It is one of the most important theorems in the theory of normed spaces. The standard proof of this theorem is based on a topological argument known as the Baire category theorem (see, for instance, E. Kreyszig (1978)) due to René-Louis Baire (1874–1932). We present a proof based on the diagonal theorem. This method was first introduced ´ by Jan Mikusinski (1970) and then extensively used by Piotr Antosik and Charles Swartz (see Antosik and Swartz (1985) and Swartz (1992)). The advantage of this approach is that it is simple and no new concepts are needed.

Theorem 1.5.12. (Diagonal theorem) Let E be a normed space and let (xij ), i, j ∈ N, be an infinite matrix of elements of E. If (a) limi→∞ xij = 0 for every j ∈ N and (b) every increasing sequence of indices (pj ) has a subsequence (qj ) such that lim

i→∞

∞ 

xqi qj = 0,

j=1

then limi→∞ xii = 0.

Proof: Suppose limi→∞ xii = 0. Then there exists an increasing sequence of indices (pi ) and some ε > 0 such that xpi pi ≥ ε for all i ∈ N. By (b), the sequence (pi ) has a subsequence (qi ) such that limi→∞ ∞ j=1 xqi qj = 0. Note that each row and each column in the matrix (xqi qj ) converges to 0. Set r1 = q1 . Now let r2 be the first index such that r2 > r1 , xqi r1 < ε/4 for all qi ≥ r2 , and xr1 r2 < ε/8. Next, let r3 be the first index such that r3 > r2 , xqi r2 < ε/8 for all qi ≥ r3 , and xrj r3 < ε/16 for j = 1, 2. In the nth step we let rn be the first index such that rn > rn−1 , xqi rn−1 < ε/2n for all qi ≥ rn , and xrj rn < ε/2n+1 for j = 1, . . . , n − 1. Continuing this process, we construct an infinite matrix (xri rj ) such that xri rj < ε/2j+1

for all i such that i = j.

In view of (b), (rj ) has a subsequence (sj ) such that limi→∞ Consider the matrix (xsi sj ). For every i ∈ N, we have

∞

j=1 xsi sj

= 0.

1.5 Linear Mappings  ∞            xsi sj  = xsi si + xsi sj     

31

i=j

j=1

        ≥  xsi si −  xsi sj    i=j

     ε  ≥  xsi si − xsi sj  ≥ . 2 i=j

This, however, is impossible since limi→∞ proves the theorem.

∞

j=1 xsi sj

= 0. This contradiction


Theorem 1.5.13. (Banach–Steinhaus theorem) Let T be a family of bounded linear mappings from a Banach space X into a normed space Y . If for every x ∈ X there exists a constant Mx such that Tx ≤ Mx for all T ∈ T , then there exists a constant M > 0 such that T ≤ M for all T ∈ T . Proof: Suppose there is no such M. Then there exist T1 , T2 , . . . ∈ T and x1 , x2 , . . . ∈ X, such that xn ≤ 1, for all n ∈ N, and Tn xn → ∞ as n → ∞. For some increasing sequence of indices (pn ), we must have Tpn xpn ≥ n2n or    1  Tp xpn  ≥ 1. (1.15)  n n 2n  xp

Consider the matrix (yij ) defined by yij = 1i Tpi 2jj , i, j ∈ N. Since the series ∞ xpj j=1 2j is absolutely convergent and X is complete, there is a z ∈ X such  xpj that z = ∞ j=1 2j . Then   ∞   ∞    1 x  pj     yij  =  Tp =     i i 2j  j=1

j=1

  ∞   x pj  1   Tp = i i 2j  j=1

 1 Tp (z) ≤ C , i i i

 where C is a constant that depends on z. Consequently limi→∞ ∞ j=1 yij = 0. Note that the same argument can be repeated for the matrix (yqi qi ) where (qi ) is an arbitrary increasing sequence of indices. Since also limi→∞ yij = 0 for all j ∈ N, the assumptions of the diagonal theorem are satisfied and we must have 1 x pi Tpi i = lim yii = 0. i→∞ i i→∞ 2 lim

But this contradicts (1.15).




32

Chapter 1

Normed Vector Spaces

1.6 Contraction Mappings and the Banach Fixed Point Theorem The name fixed point theorem is usually given to a result which says that, if a mapping f satisfies certain conditions, then there is a point z such that f (z) = z. Such a point z is called a fixed point of f . Theorems of this kind have numerous important applications. Some of them, in the theory of differential and integral equations, will be discussed in Chapter 5.

Example 1.6.1. Let E = C ([0, 1]) be the space of complex-valued continuous functions defined on the closed interval [0, 1]. Let T be defined by  t x(τ )dτ. (Tx)(t) = x(0) + 0

Clearly, for any a ∈ C, the function x(t) = aet is a fixed point of T.




Theorem 1.6.4 proved in this section is a version of a theorem called the contraction theorem or the Banach fixed point theorem. The theorem is usually formulated for metric spaces because the algebraic structure of the space is not essential.

Definition 1.6.2.

(Contraction mapping) A mapping f from a subset A of a normed space E into E is called a contraction mapping (or simply a contraction) if there exists a positive number α < 1 such that   f (x) − f (y) ≤ α x − y for all x, y ∈ A. (1.16)

Note that contraction mappings are continuous.

Example 1.6.3. Consider the nonlinear algebraic equation x3 − x − 1 = 0. This equation has three roots. There are several ways of putting the equation in the form Tx = x, for example Tx = (1 + x)1/3 ,

Tx = x3 − 1,

Tx =

x2

1 . −1

The original equation has a root in [1, 2]. The mapping T defined by T(x) = (1 + x)1/3 is a contraction on [1, 2]. Indeed, by the mean value theorem, we have   21/3 |x − y|. |Tx − Ty| = (1 + x)1/3 − (1 + y)1/3  ≤ 6

1.6 Banach Fixed Point Theorem

33

Note that the other two mappings are not contractions.




A number of important examples of contraction mappings will be discussed in Chapter 5.

Theorem 1.6.4. (Banach fixed point theorem) Let F be a closed subset of a Banach space E and let f be a contraction mapping from F into F. Then there exists a unique z ∈ F such that f (z) = z.

Proof: Let 0 < α < 1 be such that   f (x) − f (y) ≤ α x − y for all x, y ∈ F. Let x0 be an arbitrary point in F and let xn = f (xn−1 ) for n = 1, 2, . . . . We will show that (xn ) is a Cauchy sequence. First observe that, for any n ∈ N, xn+1 − xn ≤ α xn − xn−1 ≤ α 2 xn−1 − xn−2 ≤ · · · ≤ α n x1 − x0 . Hence, for any m, n ∈ N such that m < n, we have xn − xm ≤ xn − xn−1 + xn−1 − xn−2 + · · · + xm+1 − xm   ≤ α n−1 + α n−2 + · · · + α m x1 − x0 ≤

x1 − x0 m α → 0 as m → ∞. 1−α

Thus, (xn ) is a Cauchy sequence. Since F is a closed subset of a complete space, there exists a z ∈ F such that xn → z as n → ∞. We are going to show that z is the unique point such that f (z) = z. Indeed, since     f (z) − z ≤ f (z) − xn  + xn − z   = f (z) − f (xn−1 ) + xn − z ≤ α z − xn−1 + xn − z → 0 as n → ∞, we have f (z) − z = 0, and thus f (z) = z. Suppose now f (w) = w for some w ∈ F. Then   z − w = f (z) − f (w) ≤ α z − w . Since 0 < α < 1, we must have z − w = 0, which implies z = w.




Note that the proof provides a practical way of finding or approximating the fixed point. The use of this method will be discussed in Chapter 5.

Example 1.6.5. We will apply the method of successive approximation to find the solution of x3 − x − 1 = 0 in [1, 2]. In Example 1.6.3 we have shown that

34

Chapter 1

Normed Vector Spaces

Tx = (1 + x)1/3 is a contraction mapping on [1, 2] and its fixed point is the solution of the equation. We set the initial guess at x0 = 1. Then we find x1 = Tx0 ≈ 1.2599, x2 = Tx1 ≈ 1.3123, x3 = Tx2 ≈ 1.3224, x4 = Tx3 ≈ 1.3243, x5 = Tx4 ≈ 1.3246, x6 = Tx5 ≈ 1.3247, x7 = Tx6 ≈ 1.3247. Thus, the sixth iteration gives the root to four decimal places.




The number α in the definition of contraction mapping is assumed to be strictly less than 1. If this assumption is replaced by α ≤ 1 then the fixed point theorem in the above version is no longer true. Even if the condition (1.16) is replaced by strict inequality Tx − Ty < x − y

for all x, y ∈ A,

T need not have a fixed point.

Example 1.6.6. Consider the function f (x) = x + e−x as a mapping from R+

into R+ , where R+ denotes the set of all non-negative real numbers. For any x, y ∈ R+ we have, by the mean value theorem,   f (x) − f (y) < |x − y|, since |f (ξ )| < 1 for all ξ ∈ R+ . However, f is not a contraction, because there is no α < 1 such that |f (x) − f (y)| ≤ α|x − y| for all x, y ∈ R+ . Note that f does not have a fixed point.


1.7 Exercises 1. Prove that for every x, y ∈ E there exists a unique z ∈ E such that x + z = y. 2. Prove that for any vectors x, y, and z we have the following: (a) x + (y − x) = y. (b) x − (y − z) = x − y + z. 3. Let x be an element of a vector space and let λ be a scalar. Prove the following:

1.7 Exercises

35

(a) If λ = 0 and λx = 0, then x = 0. (b) If x = 0 and λx = 0, then λ = 0. (c) 0x = 0 and (−1)x = −x. 4. Check that the following are vector spaces: (a) RN . (b) CN . (c) C (), where  is a subset of RN . (d) C k (RN ). (e) C ∞ (RN ). 5. Show that the family of all solutions of the ordinary differential equation y

+ y = 0 is a vector space. 6. Show that the family of all solutions of the integral equation y(x) = b a K(x, t)y(t)dt is a vector space. 7. Show that the family of all solutions of the nonlinear ordinary differential equation y

= y2 is not a vector space. 8. Let p, q > 1 and

1 p

1

+ 1q = 1. Prove that x p ≤ 1p x +

1 q

for every 0 ≤ x ≤ 1.

9. Prove that lp is a proper vector subspace of lq whenever 1 ≤ p < q. 10. Show that any vector of R3 is a linear combination of vectors (1, 0, 0), (1, 1, 0), and (1, 1, 1). 11. Prove that every four vectors in R3 are linearly dependent. 12. Show that the functions fn (x) = xn , n = 0, 1, 2, . . . , are linearly independent. 13. Show that the functions fn (x) = enx , n = 0, 1, 2, . . . , are linearly independent. 14. Prove that spaces C (), C k (RN ), C ∞ (RN ) are infinite dimensional. 15. Denote by l0 the space of all infinite sequences of complex numbers (zn ) such that zn = 0 for all but a finite number of indices n. Find a basis of l0 . 16. According to the definition of a basis B of a vector space E, any vector x ∈ E can be represented in the form x = λ1 e1 + · · · + λn en , where e1 , . . . , en ∈ B and λ1 , . . . , λn are nonzero scalars. Prove that such a representation is unique.

36

Chapter 1

Normed Vector Spaces

17. Let E1 , . . . , En be vector spaces over the same scalar field F. If dim Ek = dk , for k = 1, . . . , n, what is the dimension of the Cartesian product E1 × · · · × En ? 18. Show that | x − y | ≤ x − y for any pair of vectors in a normed space. 19. Show that (a) (z1 , . . . , zN ) =

N

n=1 |z|,

(b) (z1 , . . . , zN ) = max{|z1 |, . . . , |zN |} are norms on CN . 20. Let  be a closed bounded subset of RN . Show that f = maxx∈ |f (x)| is a norm on C ().  p 1/p 21. Show that z = ( ∞ is a norm on lp for any p ≥ 1. n=1 |zn | ) 22. Prove that in any normed space: (a) A convergent sequence has a unique limit. (b) If xn → x and λn → λ (λn , λ are scalars), then λn xn → λx. (c) If xn → x and yn → y, then xn + yn → x + y. 23. Show that two norms are equivalent if and only if they define the same bounded sets. 24. Using Theorem 1.3.11 prove that the following norms on R2 are equivalent:     (x, y) = x2 + y2 , (x, y) = |x| + |y|, 1 2  
 (x, y) = max |x|, |y| . 3 25. Prove Theorem 1.3.20. 26. Consider the convergence in C (R) defined as follows: fn → f if the sequence (fn ) converges to f uniformly on every bounded interval [a, b]. Is it possible to define this convergence by a norm on C (R) ? 27. Show that the only dense linear subspace of CN is CN itself. 28. Find a sequence of elements of l1 that is convergent in l2 but divergent in l1 . What about the converse? Can you give similar example for any pair of spaces lp and lq with p = q ? 29. Show, without using Theorem 1.3.13, that the following norms on CN are equivalent:   (z1 , . . . , zN ) = |z1 |2 + · · · + |zN |2 ,

1.7 Exercises

37

N    (z1 , . . . , zN ) = |zn |, n=1

 
 (z1 , . . . , zN ) = max |z1 |, . . . , |zN | . 30. Find two norms on C ([0, 1]) which are not equivalent. 31. Prove that a subset S of a normed space is bounded if and only if λn xn → 0 for any xn ∈ S and any scalars λn → 0. 32. Show that, in a normed space, xn → x implies n1 (x1 + · · · + xn ) → x. 33. Show that, in a normed space, if every subsequence of a sequence (xn ) contains a subsequence convergent to x, then (xn ) itself converges to x. 34. Prove Theorem 1.3.27. 35. Let E1 and E2 be normed spaces. Show that a function f : E1 → E2 is continuous if and only if for every x ∈ E1 and ε > 0 there exists a δ > 0 such that f (x) − f (y) < ε whenever x − y < δ. 36. Let E1 and E2 be normed spaces and let f : E1 → E2 be a continuous function. Prove that, if S ⊂ E1 is compact, then f (S) is compact in E2 . 37. Let f : E → R be a continuous function and let S ⊂ E be compact. Prove that f attains the minimum and the maximum values on S. 38. Is the sequence of functions fn (x) = C ([0, 1]) ?

nx 1+nx2

a Cauchy sequence in

39. Prove Theorem 1.4.2. 40. Show that the sum of two Cauchy sequences is a Cauchy sequence. 41. Prove completeness of lp for arbitrary p ≥ 1. 42. Give an example of an incomplete normed space. 43. Show that, in a Banach space, every sequence (x n ) convergent to zero contains a subsequence (xpn ) such that the series ∞ n=1 xpn converges. b 44. Consider the space C ([a, b]) with the norm defined as f = a |f (t)|dt. Is this a Banach space ? x 45. Show that L(f )(x) = 0 f (t)dt defines a continuous linear mapping from C ([0, 1]) into itself. 46. Give an example of a linear mapping from a normed space into a normed space which is not continuous. 47. Prove that a linear mapping from a normed space into a normed space is continuous if and only if it maps bounded sets to bounded sets.

38

Chapter 1

Normed Vector Spaces

48. Prove Theorem 1.5.4. 49. Prove Theorem 1.5.11. 50. Show that B (E1 , E2 ) is a vector space. 51. A sequence of mappings Ln ∈ B (E1 , E2 ) converges strongly to L ∈ B(E1 , E2 ) if for every x ∈ E1 we have Ln x − Lx → 0 as n → ∞. Show that uniform convergence implies strong convergence, but not conversely. 52. Let E = C ∞ ([a, b]) be the space of all infinitely differentiable functions on the interval [a, b] with f = max[a,b] |f (x)|. Is the differential operd a contraction mapping ? ator D = dx 53. Define T : R2 → R2 by T(x, y) = (y1/3 , x1/3 ). What are the fixed points of T ? In which quadrants of the xy-plane is T a contraction ? 54. Is the mapping T : R → R defined by  x − 12 ex Tx = − 12 + 12 x

for x ≤ 0, for x > 0

a contraction ? 55. Consider the space RN . Elements of RN are denoted by x = (x1 , . . . , xN ), y = (y1 , . . . , yN ), and so on. The norm on RN is defined by x − y = max1≤n≤N |xn − yn |. Let T : RN → RN defined by Tx = y where yk =  N n=0 akn xn + bk , k = 1, . . . , N. Under what conditions is T a contraction mapping ?

The Lebesgue Integral CHAPTER

2

“In contrast to the classical Riemann integral, my new concept of the integral provides a better answer to the question of the connection between differentiation and integration (main theorem of calculus).” Henri Leon Lebesgue

2.1 Introduction The main purpose of this book is to present basic methods and applications of Hilbert spaces. One of the most important examples of Hilbert spaces, from the point of view of both theory and applications, is the space of Lebesgue square integrable functions on RN . Thus, the Lebesgue integral is essential for understanding some of the most important aspects of Hilbert space theory. Since Henri Lebesgue (1875–1941), in a series of five short notes published between 1899 and 1901, introduced a new integral, now universally called the Lebesgue integral, numerous approaches to the theory have been developed. In the most common approach, one studies first the concept of measure and then defines the integral. This is not a very efficient way if what is needed is the integral, not measure. Moreover, the theory of measure is often considered too difficult to be taught in an undergraduate course. In this book, we introduce the Lebesgue integral following the idea of ´ (1913–1987). This approach is H.M. MacNeille (1907–1973) and J. Mikusinski fast and direct (integral is introduced without auxiliary concepts like measure) and can be easily taught at an undergraduate level. Since the case of real-valued functions on R exhibits the essential ideas, we first present the construction in great detail in that case. Then we generalize it to complex-valued functions

39

40

Chapter 2

The Lebesgue Integral

on R. Finally, we describe the extension of the construction to complex-valued functions on RN .

2.2 Step Functions By a step function on the real line R we mean a finite linear combination of characteristic functions of semiopen intervals [a, b) ⊆ R. Thus, for every step function f , there are intervals [a1 , b1 ), . . . , [an , bn ) and numbers λ1 , . . . , λn ∈ R such that f = λ1 f 1 + · · · + λ n f n ,

(2.1)

where fk is the characteristic function of [ak , bk ), that is, fk (x) = 1 if x ∈ [ak , bk ), and f (x) = 0 otherwise. Clearly, representation of a step function in the form (2.1) is not unique. On the other hand, if we assume that intervals [ak , bk ) are disjoint and the minimal number of intervals is used, then the representation is unique. Such a representation can be obtained in the following way: Let f be a step function and let a0 , a1 , . . . , an be all points of discontinuity of f . In other words, a0 , a1 , . . . , an are the points where the graph of f has a jump. We can always assume that the points are indexed so that a0 < a1 < · · · < an . Denote by gk (k = 1, . . . , n) the characteristic function of the interval [ak−1 , ak ). Then f = α1 g1 + · · · + αn gn , where αk = f (ak−1 ), k = 1, . . . , n. This representation satisfies the required conditions. It is called the basic representation of f . Since this definition does not make much sense if f = 0, we call f = 0 the basic representation of the zero function. The collection of all step functions on R is a vector space. The absolute value of a step function is again a step function. If f = α1 f1 + · · · + αn fn is the basic representation of a step function f , then |f | = |α1 |f1 + · · · + |αn |fn . For any real-valued functions f and g, we have min{f , g} =

 1 f + g − |f − g| 2

and

max{ f , g} =

 1 f + g + |f − g| . 2

Figure 2.1 A “typical” step function.

2.2 Step Functions

41

Thus, if f and g are step functions, then min{f , g} and max{f , g} are also step functions. Finally, it is easy to see that the collection of all step functions is closed under translations, that is, if f is a step function and τz f is the function defined by τz f (x) = f (x − z) for some z ∈ R, then τz f is a step function. Note that if f is the characteristic function of [a, b), then τz f is the characteristic function of [a + z, b + z). Hence, if f = λ1 f1 + · · · + λn fn is the basic representation of f , then τz f = λ1 τz f1 + · · · + λn τz fn is the basic representation of τz f . By the support of a nonzero function f , denoted by supp f , we mean the set of all points x ∈ R for which f (x) = 0. The support of a nonzero step function is always a finite union of semiopen intervals. If f = 0 , supp f = ∅.

Definition 2.2.1. (Integral of a step function)  The integral f of a step function f (x) = λ1 f1 (x) + · · · + λn fn (x), where fk is the characteristic function of [ak , bk ), k = 1, . . . , n, is defined by  f = λ1 (b1 − a1 ) + · · · + λn (bn − an ).

 Clearly, the value f is equal to the Riemann integral of f . From the properties of the Riemann integral it follows that the defined integral does not depend on a particular representation. This fact is of importance for construction of the Lebesgue integral. The independence can be proved without using properties of the Riemann integral. The reader is asked to provide an elementary proof as an exercise.

Theorem 2.2.2. For any step functions f and g we have (a) (b) (c) (d) (e)



  ( f + g) = f + g;   λf = λ f , λ ∈ R;   f ≤ g implies f ≤ g;   | f | ≤ |f |;   τz f = f , z ∈ R.

Proof: Properties (a) and (b) follow directly from Definition 2.2.1. To   prove

(c) we first show that f ≥ 0 implies f ≥ 0. Indeed, if f = 0, then f = 0 by (b). If f ≥ 0 and f does not vanish identically on R, then allthe coefficients in the basic representation of f are  positive, and thus f > 0. Now, if f ≤ g, then g − f ≥ 0 and hence (g − f ) ≥ 0, which    gives  us f ≤ g, by (a) and (b). Since f ≤ |f | and −f ≤ |f |, we have f ≤ |f |

42

Chapter 2

The Lebesgue Integral

    and (−f ) ≤ |f |, by (c), which implies | f | ≤ |f |, by (b). Property (e) is obvious in view of the earlier remarks.
A rather obvious property of the integral of step functions is formulated in the following lemma. It will be used in the proof of Theorem 2.2.6. The easy proof is left as an exercise.

Lemma 2.2.3. Let f be a step function whose support is contained in the union of semiopen intervals [a1 , b1 ), . . . , [an , bn ). If |f | < M, for some constant M, then  |f | ≤ M

n  (bk − ak ). k=1

Lemma 2.2.4. Let [a1 , b1 ), [a2 , b2 ), . . . be a partition of an interval [a, b), that is, the intervals [a1 , b1 ), [a2 , b2 ), . . . are disjoint and ∞ 

[an , bn ) = [a, b).

(2.2)

n=1

Then ∞  (bn − an ) = b − a.

(2.3)

n=1

Proof: If c ∈ (a, b], then [a1 , b1 ) ∩ [a, c), [a2 , b2 ) ∩ [a, c), . . . is a partition of [a, c). Denote by S the set of all points c ∈ (a, b] such that  (bc,n − an ) = c − a,

(2.4)

an 0 and let n0 ∈ N be such that ∞  

|fn | < ε,

n=n0 +1

(n0 exists by (a) in Definition 2.3.1). Define gn = f1 + · · · + fn0 + |fn0 +1 | + · · · + |fn0 +n |,

(2.11)

46

Chapter 2

The Lebesgue Integral

for n = 1, 2, . . . . Since (gn ) is a nondecreasing sequence of step  functions such that limn→∞ gn (x) ≥ 0 for all x ∈ R, we have limn→∞ gn ≥ 0, by Corollary 2.2.7. Thus,  0≤

 f1 + · · · +

fn0 +

∞  

|fn |

n=n0 +1

 ≤

 f1 + · · · +

fn0 +

∞  

 fn +

 ∞   |fn | + |fn |

n=n0 +1

=

∞  

fn + 2

∞  

n=n0 +1

 |fn |

n=n0 +1

n=1

≤

∞ 

fn + 2ε,

n=1

proving the lemma, since ε is arbitrary.
   ∞ ∞ ∞ Corollary 2.3.3. If f  ∞ n=1 fn and f  n=1 gn , then n=1 fn = n=1 gn .

Proof: Since 0  f1 − g1 + f2 − g2 + · · · , we have, by Lemma 2.3.2, 

 f1 − and hence

 g1 +

∞  

fn −

n=1

 f2 − ∞  

g2 + · · · ≥ 0

gn ≥ 0,

n=1

because both series are absolutely convergent. Similarly, we obtain ∞   n=1

gn −

∞  

fn ≥ 0,

n=1

and therefore the sums of both series are equal.




Corollary 2.3.3 implies that step functions are integrable and that in this case integrals in Definition 2.2.1 and Definition 2.3.1 are equal (which justifies the use of the same symbol in both cases). Indeed, if f is a step function, then f  f + 0 + 0 + · · · . It can be proved (see Section 2.10) that every Riemann integrable function is Lebesgue integrable and both integrals are equal. The space of all Lebesgue integrable functions defined on R is denoted by L1 (R). In the remainder of this book, Lebesgue integrable functions are called simply integrable.

2.3 Lebesgue Integrable Functions

Theorem 2.3.4. L1 (R) is a vector space and 

47



1 is a linear functional on L (R). f ≤ g.

Moreover, if f , g ∈ L (R) and f ≤ g, then  ∞ Proof: If f  ∞ n=1 fn , g  n=1 gn , and λ ∈ R, then 1

f + g  f1 + g1 + f2 + g2 + · · · and λf  λf1 + λf2 + · · · , and consequently f + g ∈ L1 (R) and λf ∈ L1 (R). Moreover,      (f + g) = f + g and λf = λ f .  If f ≤g, then  g − f ≥ 0, and hence (g − f ) ≥ 0, by Lemma 2.3.2. This implies g − f ≥ 0, proving the theorem.
Definition 2.3.1 is fairly simple, although it may be a little disturbing. Condition (a) does not involve f in any way. Condition (b) says that f equals the sum of series ∞ n=1 fn at those points where the series converges absolutely. How big is the set?  If we could find a sequence of step functions fn satisfying (a) such that ∞ n=1 |fn | diverges at every point of an open interval (a, b), then we could write f  f1 + f2 + · · · for any function f that vanishes outside of (a, b). Since we have proved that the integral is well defined, we know that such a sequence (fn ) cannot exists. In fact, the series cannot diverge at every point of a nonempty open interval. Actually, we can prove  more: If f  f1 + f2 + · · · and Z denotes the set of all points where the series ∞ n=1 |fn (x)| diverges, then the integral of the characteristic function of Z equals 0. Indeed, if g is the characteristic function of Z, then f + g  f1 + f2 + · · · , and thus





(f + g) = f .  This,in view of Theorem 2.3.4, implies g = 0. If g = 0, then the set Z must be, in some  sense, a small set. Such a set is called a null set. We can say that series ∞ n=1 fn (x) converges to f (x) for all x except a null set. This type of convergence is called convergence almost everywhere. Null sets and convergence almost everywhere will be discussed in Section 2.7. At this point the reader may expect some examples of Lebesgue integrable functions. We know that every step function is Lebesgue integrable, but this is definitely not a satisfactory example. For instance, we would like to see examples of functions which are Lebesgue integrable but not Riemann integrable.

48

Chapter 2

The Lebesgue Integral

Unfortunately all we have at this time is the definition of the integral, and it would be awkward to try to prove that a function is integrable by expanding the function into a series of step functions. It would be like trying to evaluate 1 x 0 e dx using Riemann sums. The reader has to be patient and let the theory develop a little bit further, and then we will easily construct nontrivial examples of Lebesgue integrable functions.

2.4 The Absolute Value of an Integrable Function In the previous section, we mention that all Riemann integrable functions are Lebesgue integrable. (Recall that, by definition, a Riemann integrable function vanishes outside a bounded interval.) For improper integrals it is not so. For instance, the integral  ∞ sin x dx −∞ x converges, although the function is not Lebesgue integrable, as we can see from the next theorem.

Theorem 2.4.1. If f ∈ L1 (R) and f  f1 + f2 + · · · , then |f | ∈ L1 (R) and we have    

  ∞    f  ≤ |f | ≤ |fn |. n=1

Proof: Let f  f1 + f2 + · · · . Define

 ∞    fn (x) < ∞ Z = x ∈ R: 

n=1

and sn = f1 + · · · + fn ,

for n = 1, 2, . . . .

Then f (x) = limn→∞ sn (x) for every x ∈ Z. Hence also     f (x) = lim sn (x), n→∞

or, equivalently,             f (x) = s1 (x) + s2 (x) − s1 (x) + s3 (x) − s2 (x) + · · · for every x ∈ Z. If we let g1 = |s1 | and gn = |sn | − |sn−1 | for n ≥ 2, then ∞    f (x) = gn (x) n=1

2.4 The Absolute Value of an Integrable Function

49

for every x ∈ Z. Moreover, since   |gn | = |sn | − |sn−1 | ≤ |sn − sn−1 | = |fn |, (2.12)     we have |gn | ≤ |fn |, and hence ∞ n=1 |gn | < ∞. It may seem that |f |  g + g + · · · , but there may exist points not in Z  for which the series 1 2 ∞ ∞ |g (x)| converges and we cannot guarantee that n n=1 n=1 gn (x) = |f (x)| at those points. Wecan get rid of those points by adding and subtracting terms of the series ∞ n=1 fn : |f |  g1 + f1 − f1 + g2 + f2 − f2 + · · · .

(2.13)

This does not change convergence at points of Z, but makes the series diverge at other points. Since       |g1 | + |f1 | + |f1 | + |g2 | + |f2 | + |f2 | + · · · < ∞, the expansion (2.13) is valid and thus |f | is integrable. Moreover,        |f | = g1 + f1 − f1 + g2 + f2 − f2 + · · · 

 g1 +

=

g2 + · · · ≤

 |g1 | +

|g2 | + · · ·



 ≤



|f1 | +

|f2 | + · · · .

Finally, since   f ≤ |f | and −f ≤ |f |, we have and hence  f  ≤ |f |.



f≤



  |f | and − f ≤ |f |,


Corollary 2.4.2. If f , g ∈ L1 (R), then min{f , g}, max{f , g} ∈ L1 (R). Proof: Since min{f , g} =

 1 f + g − |f − g| 2

and

max{f , g} =

the assertion follows by Theorems 2.3.4 and 2.4.1.

 1 f + g + |f − g| , 2


Integrability of |f | does not imply integrability of f . However, it is not possible to present an example of such a function. We can only prove that such a function exists. The proof of existence requires mathematical tools that will not be introduced in this book. The interested reader should refer to books on real analysis or measure theory (see, for example, Friedman (1982) or Halmos (1974)). The following theorem establishes a useful property of translations of integrable functions (see, for example, the proof of Theorem 2.15.3). Note that if

50

Chapter 2

The Lebesgue Integral

then the translated function τz f is integrable for any z ∈ R and f is integrable,  τz f = f .  Theorem 2.4.3. If f ∈ L1 (R), then limz→0 |τz f − f | = 0.

Proof: Note that the proof is easy if f is a step function. Now let f be an arbitrary integrable function and let ε > 0. If f  f1 + f2 + · · · , then there exists an n0 ∈ N such that ∞   n=n0 +1

ε |fn | < . 3

We have 

    n0 n0 ∞  ∞        |τz f − f | ≤  τ z fn − fn  + |τz fn | + |fn |   n=1 n=1 n=n0 +1 n=n0 +1     n0 n0 ∞       =  τ z fn − fn  + 2 |fn |   n=1 n=1 n=n0 +1     n0 n0  2ε    τ z fn − fn  + . <    3 n=1 n=1

Since

n0

n=1 fn

is a step function, we have     n0 n0     τ z fn − fn  = 0. lim  z→0   n=1

Consequently,



n=1

|τz f − f | < ε for sufficiently small z.




2.5 Series of Integrable Functions In the construction described in the preceding sections we start with an integral defined for step functions and extend it to a larger class of functions by expending them in a series of step functions f  f1 + f2 + · · · . Are we going to obtain new functions if we repeat the construction and use integrable functions instead of step functions? It turns out that we do not get anything new. This result is a form of completeness of the space of Lebesgue integrable functions. It is of fundamental importance for the theory of the Lebesgue integral. First we extend the use of the notation f  f1 + f2 + · · · to a series of integrable functions.

51

2.5 Series of Integrable Functions

Definition 2.5.1. Let f be a real-valued function and let (fn ) be a sequence of integrable functions. If (a)

∞  

|fn | < ∞,

n=1

(b)

f (x) =

∞ 

for every x ∈ R such that

fn (x)

n=1

∞ 

|fn (x)| < ∞,

n=1

then we write f  f1 + f2 + · · · or f 

∞

n=1 fn .

The following technical lemma will be used in the proof of the important Theorem 2.5.3.

Lemma 2.5.2. If f ∈ L1 (R), then for every ε  > 0 there  exists a sequence of step ∞ n=1

functions (fn ) such that f  f1 + f2 + · · · and

|fn | ≤ |f | + ε.

Proof: Let f  g1 + g2 + · · · be an arbitrary expansion f in a series of of ∞ ε

step functions. Then there exists an n0 ∈ N such that Define f1 = g1 + · · · + gn0

n=n0 +1

|gn | < 2 .

fn = gn0 +n−1 for n ≥ 2.    Then obviously f  f1 + f2 + · · · . Since |f1 | − |f | ≤ |f1 − f | and f − f1  f2 + f3 + · · ·, we get and



 |f1 | −

|f | ≤

∞  

|fn |

n=2

and hence

 |f1 | −

∞  

 |fn | ≤

|f |.

n=2

Consequently, ∞  

 |fn | =

|f1 | +

n=1

∞  

|fn |

n=2

 =

|f1 | −

∞  

|fn | + 2

n=2

 ≤

|f | + 2

∞   n=2

∞   n=2

|fn |

|fn |

52

Chapter 2

The Lebesgue Integral

 =

∞  

|f | + 2 

|gn |

n=n0 +1

|f | + ε.


0. There exists an integer n0 such that

∞ n=n0



|fn | < ε. Since

f − f1 − · · · − fn  fn+1 + fn+2 + · · · , for every n > n0 , we have    |f − f1 − · · · − fn | ≤ |fn+1 | + |fn+2 | + · · · < ε, by Theorem 2.4.1.




2.7 Convergence Almost Everywhere

55

2.7 Convergence Almost Everywhere If f  f1 + f2 + · · · , then the series f1 (x) + f2 (x) + · · · need not converge at every x ∈ R. On the other hand, the set of those points where we do not have convergence is a “small set.” This suggests a new type of convergence, which is very natural when the Lebesgue integral is considered.

Definition 2.7.1. (Null set) A set X ⊆ R is called a null set or a set of measure zero if its characteristic function is a null function. Every countable set is a null set. A countable union of null sets is a null set. There are null sets which are not countable, for example, the Cantor set (see Exercise 20). The following theorem is an immediate consequence of Theorem 2.6.3.

Theorem 2.7.2. Every subset of a null set is a null set. Definition 2.7.3.

(Equality almost everywhere) Let f and g be functions defined on R. If the set of all x ∈ R for which f (x) = g(x) is a null set, then we say that f equals g almost everywhere and write f = g a.e. 

Theorem 2.7.4. f = g a.e. if and only if |f − g| = 0. Proof: Let h be the characteristic function of the set Z of all x ∈ R for which f (x) = g(x).   If f = g a.e., then |h| = h = 0. Therefore |f − g|  h + h + · · · , 

which implies |f − g| = 0. Conversely, if |f − g| = 0, then h  |f − g| + |f − g| + · · · , and hence



h = 0. This shows that Z is a null set, that is, f = g a.e.




Note that Theorem 2.7.4 implies that f is a null function if and only if f = 0 a.e. Similarly, the equivalence of functions defined in Section 2.6 and equality almost everywhere are the same notion. This important property is used often in arguments concerning the Lebesgue integral. Another useful consequence of Theorem 2.7.4 is that we do not need to know the value of a function at

56

Chapter 2

The Lebesgue Integral

every point in order to find its integral. It is sufficient to know the values almost everywhere, that is, everywhere except a null set. The function need not even be defined at every point.

Theorem 2.7.5. Suppose fn → f i.n. Then fn → g i.n. if and only if f = g a.e. Proof: If fn → f i.n. and f = g a.e., then 

fn − g =



|fn − g| ≤



|fn − f | +

|f − g| =

 |fn − f | = fn − f → 0.

If fn → f i.n. and fn → g i.n., then    |f − g| ≤ |f − fn | + |fn − g| → 0, completing the proof.




Definition 2.7.6.

(Convergence almost everywhere) We say that a sequence of functions f1 , f2 , . . . defined on R converges to f almost everywhere, denoted by fn → f a.e., if fn (x) → f (x) for every x except a null set.

Convergence almost everywhere has properties similar to convergence in norm. If fn → f a.e. and λ ∈ R, then λfn → λf a.e. If fn → f a.e. and gn → g a.e., then fn + gn → f + g a.e. If fn → f a.e., then |fn | → |f | a.e.

Theorem 2.7.7. Suppose fn → f a.e. Then fn → g a.e. if and only if f = g a.e. Proof: If fn → f a.e. and fn → g a.e., then fn − fn → f − g a.e., which means that f − g = 0 a.e. Now, assume fn → f a.e. and f = g a.e. Denote by A the set of all x ∈ R such that the sequence (fn (x)) does not converge to f (x) and by B the set of all x ∈ R such that f (x) = g(x). Then A and B are null sets and so is A ∪ B. Since fn (x) → g(x) for every x not in A ∪ B, we have fn → g a.e.
The following two examples show that convergence in norm and convergence almost everywhere are essentially different.

Example 2.7.8. For n = 1, 2, . . . , define    √1 n fn (x) =  0

for x ∈ [−n, n], otherwise.

2.7 Convergence Almost Everywhere

57

Then fn (x) → 0 for every x ∈ R, and thus fn → 0 a.e. On the other hand, since  √ |fn | = 2 n → ∞, the sequence is not convergent in norm.




Example 2.7.9. Consider the sequence (fn ) of functions defined as follows:  fn (x) =

1 0

if 2k−1 ≤ n < 2k and x ∈ otherwise,



n 2k−1

 − 1, 2n+1 k−1 − 1 ,

n = 1, 2, . . . (see Figure 2.2). Since  1 fn = k−1 for 2k−1 ≤ n < 2k , 2   |fn | = fn → 0, and thus the sequence (fn ) converges to 0 in norm. On the other hand, for every x ∈ [0, 1), fn (x) = 0 for infinitely many n ∈ N and fn (x) = 1 for infinitely many n ∈ N. Therefore, the sequence (fn ) does not converge almost everywhere.
Example 2.7.8 can be used to show that, in general, even uniform convergence does not imply convergence in norm. However, if we assume that f , f1 , f2 , . . . are integrable functions that vanish outside a bounded interval [a, b] and the sequence (fn ) converges to f uniformly, then fn → f i.n. Indeed,

Figure 2.2 Functions f1 , . . . , f8 in Example 2.7.9.

58

Chapter 2

in this case we have 

The Lebesgue Integral

  |fn − f | ≤ (b − a) sup fn (x) − f (x) → 0. [a,b]





Theorem 2.7.10. Let f1 , f2 , . . . ∈ L1 (R), and |f1 | + |f2 | + · · · < ∞. Then the

series f1 + f2 + · · · converges almost everywhere.

Proof: By Corollary 2.5.4,  there exists a function f ∈ L1 (R) such that f 

∞ f1 + f2 + · · · . Since f (x) = ∞ n=1 fn (x) for every x such that n=1 |fn (x)| < ∞, it suffices to show that the set of all points x ∈ R for which the series ∞ |f (x)| is not absolutely convergent is a null set. Let g be the characn=1 n teristic function of that set. Then g  f1 −f1 +f2 −f2 +· · · , and consequently       |g| = g = f1 − f1 + f2 − f2 + · · · = 0.


Theorem 2.7.10 leads us to an important corollary.

Corollary 2.7.11. If f  f1 + f2 + · · · , then f = f1 + f2 + · · · a.e.  |f1 | + |f2 | + · · · < ∞. Then f = f1 + f2 + · · · a.e. if and only if f = f1 + f2 + · · · i.n.

Theorem 2.7.12. Let f1 , f2 , . . . ∈ L1 (R), and



Proof: By Corollary 2.5.4, there exists a function g ∈ L1 (R) such that g 

f1 + f2 + · · · . Then, by Theorem 2.6.6 we have g = f1 + f2 + · · · i.n., and by Corollary 2.7.11 we have g = f1 + f2 + · · · a.e. Now, if f = f1 + f2 + · · · a.e., then f = g a.e., by Theorem 2.7.7. Hence f = f1 + f2 + · · · i.n., by Theorem 2.7.5. Conversely, if f = f1 + f2 + · · · i.n., then f = g a.e., by Theorem 2.7.5. Hence f = f1 + f2 + · · · a.e., by Theorem 2.7.7.


2.8 Fundamental Convergence Theorems Now we are ready to prove the main theorems of this chapter. The first theorem implies that L1 (R) (or, more precisely, L1 (R), defined in Section 2.6) is a Banach space.

Theorem 2.8.1. The space L1 (R) is complete. Proof: In view of Theorem 1.4.9, it suffices to prove that every absolutely convergent series converges in norm. Let fn ∈ L1 (R), n = 1, 2, . . . , and ∞  1 n | < ∞. Then, by Corollary 2.5.4, there exists an f ∈ L (R) such n=1 |f ∞ that ∞ f  n=1 fn . This in turn implies, by Theorem 2.6.6, that the series
n=1 fn converges to f in norm, proving the theorem.

2.8 Fundamental Convergence Theorems

59

In view of the previous theorem, the space L1 (R) can be defined as the completion of the space of step functions with respect to convergence in norm. This looks like a very simple definition of the Lebesgue integrable functions. However, in such an approach elements of L1 (R) are not functions and it is not at all clear if and how we can associate a function with an element of L1 (R). Moreover, the important relations between convergence in norm and convergence almost everywhere do not follow easily from such a definition. We know that convergence in norm and convergence almost everywhere are essentially different. The next theorem connects both convergences in an elegant way.

Theorem 2.8.2. If fn → f i.n., then there exists a subsequence (fpn ) of (fn ) such

that fpn → f a.e.



Proof: Since |fn − f | → 0, there exists an increasing sequence of positive

integers (pn ) such that

|fpn − f | < 2−n . Then 

 |fpn+1 − fpn | ≤

 |fpn+1 − f | +

|f − fpn |
0, then |fn g| < M |fn | for all n ∈ N. Thus, ∞  

|fn g| < M

n=1

∞  

|fn | < ∞.

n=1

∞  Moreover, if ∞ n=1 |g(x)fn (x)| < ∞ for some ∞ x ∈ R, then n=1 |fn (x)| < ∞ or g(x) = 0. In either case, F(x)G(x) = n=1 fn (x)g(x) for that x.


Theorem 2.9.5. If f is a locally integrable function such that |f | ≤ g for some g ∈ L1 (R), then f ∈ L1 (R). Proof: Let fn = f χ[−n,n] for n = 1, 2, . . . . Then f1 , f2 , . . . are integrable functions, the sequence (fn ) converges to f everywhere and |fn | ≤ g for every n ∈ N. Thus, by the Lebesgue dominated convergence theorem, f ∈ L1 (R).
b In applications it often convenient to use the symbol a f even if b ≤ a. We adopt the usual convention:  a  a  b f =− f and f = 0. a

b

a

The subspace of L1 (R) of all functions vanishing outside of some interval [a, b] is a vector space. It is denoted by L1 ([a, b]). As a subspace of L1 (R), it is a normed space (with functions equal almost everywhere identified). The norm in L1 ([a, b]) is then defined by  b f = |f |. a

64

Chapter 2

The Lebesgue Integral

Completeness of L1 ([a, b]) follows easily from completeness of L1 (R). Clearly, all properties of the Lebesgue integral apply to functions vanishing outside of some interval. Instead of intervals other sets can be used as well. This will be discussed further in Section 2.11.

2.10 The Lebesgue Integral and the Riemann Integral In this section we are going to prove that Riemann integrable functions (Georg Friedrich Bernhard Riemann (1826–1866)) are Lebesgue integrable. Then we will prove some theorems, which show that useful properties of the Riemann integral are not lost when the more general Lebesgue integral is considered.

Theorem 2.10.1. A Riemann integrable function is Lebesgue integrable and the integrals are equal. Proof: Let f be a Riemann integrable function vanishing outside an interval [a, b) and let |f | < M for some constant M. For every n ∈ N, we partition the interval [a, b) into n intervals of equal length, [a, a + c), [a + c, a + 2c), . . . , [a + (n − 1)c, a + nc), where c = (b − a)/n. Then we define functions 
 inf f (t): t ∈ [a + (k − 1)c, a + kc) if x ∈ [a + (k − 1)c, a + kc), gn (x) = 0 otherwise, and  hn (x) =


 sup f (t): t ∈ [a + (k − 1)c, a + kc) 0

if x ∈ [a + (k − 1)c, a + kc), otherwise.

Since (gn ) is a bounded nondecreasing sequence of functions, it converges at every x ∈ R. Define g(x) = limn→∞ gn (x). Functions gn are integrable, being step functions, and thus, by the Lebesgue dominated convergence theorem, g is integrable and   g = lim gn . n→∞

Similarly, the function defined by h(x) = limn→∞ hn (x) is integrable and   hn . h = lim n→∞

2.10 The Lebesgue Integral and the Riemann Integral

65

On the other hand, since f is Riemann integrable, we have   lim gn = lim hn , n→∞

and thus



n→∞

 |h − g| =

 (hn − gn ) = 0.

h − g = lim

n→∞

(2.15)

Hence, g = h a.e., and since g ≤ f ≤ h, we have g = f a.e. (as well as h = f a.e.). This proves that f is Lebesgue integrable. Clearly, the Riemann integral off and the Lebesgue integral of f are equal, both being equal to limn→∞ gn .


Theorem 2.10.2. Let f be a Lebesgue integrable function on a bounded  x or unbounded interval (a, b), and let a ≤ c ≤ b. Then the function F(x) = c f is continuous in (a, b) and the limits F(a) = lim F(x) and

F(b) = lim F(x)

x→a+

x→b−

exist and are finite.

Proof: Let x be a point such that a ≤ x < b, and let (xn ) be a decreasing sequence in (a, b) convergent to x, that is, x < xn+1 < xn < b and xn → x. Let gn = χ(x,xn ) , n = 1, 2, . . . . Then the sequence of products fgn is convergent to zero at every point of (a, b). Since |fgn | ≤ |f |, we have 

xn

F(xn ) − F(x) =



b

f=

x

fgn → 0,

a

by the Lebesgue dominated convergence theorem. Thus, F is right-hand continuous at x ∈ (a, b) and has a limit at a. Similarly, we can prove lefthand continuity at every x, such that a < x ≤ b, by considering the characteristic functions of intervals (xn , x) where a < xn < xn+1 < x and xn → x. Since F is right-hand and left-hand continuous, it is continuous in (a, b) and the limits F(a) and F(b) exist.
 Theorem 2.10.3. Let f be a function integrable over (a, b) and let F(x) = cx f for some c ∈ (a, b). If f is continuous at a point x0 ∈ (a, b), then F is differentiable at x0 and we have F (x0 ) = f (x0 ).

Proof: For h > 0, we have F(x0 + h) − F(x0 ) = h

 x0 +h c

f− h

 x0 c

f

1 = h



x0 +h

f. x0

(2.16)

66

Chapter 2

The Lebesgue Integral

Since f is continuous at x0 , for an arbitrary ε > 0 there is a δ > 0 such that f (x0 ) − ε < f (t) < f (x0 ) + ε

for all t ∈ (x0 , x0 + δ),

Integrating (2.17) from x0 to x0 + h, we get  x0 +h     f < h f (x0 ) + ε , h f (x0 ) − ε
0. By Lemma 2.5.2, there a sequence of step functions (fn )  ∞ exists such that f  f1 + f2 + · · · and n=1 |fn | < ε/3. For all n ∈ N define      1 . An = x ∈ R: f1 (x) + · · · + fn (x) ≥ 2 Note that these sets are finite unions of intervals because |f1 | + · · · + |fn | is a step function for every n ∈ N. Let An = [an,1 , bn,1 ) ∪ · · · ∪ [an,kn , bn,kn ). n Then kk=1

(bn,k − an,k ) < 2ε/3 for every n ∈ N. Since A1 ⊆ A2 ⊆ · · · and S ⊆ ∞ n=1 An , we can find disjoint intervals [a1 , b1 ), [a2 , b2 ), . . . satisfying (2.21). Now assume that, for every ε > 0, there exist intervals [a1 , b1 ), [a2 , b2 ), . . . satisfying (2.21). Without loss of generality, we can assume that the intervals are disjoint. Let, for every n ∈ N, [an,1 , bn,1 ), [an,2 , bn,2 ), . . .

2.11 Lebesgue Measure on R

69

be a sequence of disjoint intervals such that S⊆

∞  [an,k , bn,k )

(2.22)

k=1

and ∞  1 (bn,k − an,k ) < . n

(2.23)

k=1

Denote by gn,k the characteristic function of the interval [an,k , bn,k ), and define fn (x) =

∞ 

gn,k (x).

k=1

Note that for every n ∈ N, fn is a well-defined integrable function and  fn < n1 . Consider the sequence of functions hn = min{f1 , f2 , . . . , fn }. Since this is a nonincreasing sequence of non-negative functions, the limit h(x) = limn→∞ hn (x) exists for every x ∈ R and h is integrable, by the monotone convergence theorem. Moreover, hn ≤ fn < n1 implies that h is a null function. Let g denote the characteristic function of the set S. By (2.22), g ≤ fn , and hence g ≤ hn , for all n ∈ N. Thus, g ≤ h and, by Theorem 2.6.3, g is a null function. This means that S is a null set.
The property of measure proved in the next theorem is called σ -additivity. It is one of the fundamental properties of the Lebesgue measure.

Theorem 2.11.4.

Let S1 , S2 , . . . be a sequence of disjoint measurable sets. Then

the union S =

∞ n=1 Sn

is measurable and µ(S) =

∞ 

(2.24)

µ(Sn ).

n=1

Proof: Suppose first that there exists a bounded interval [a, b] such that Sn ⊆ [a, b] for every n ∈ N. Then χS  χS1 + χS2 + · · · . Consequently, S is measurable and we have  µ(S) =

χS =

∞   n=1

χSn =

∞  n=1

µ(Sn ).

(2.25)

70

Chapter 2

The Lebesgue Integral

Now let S, S1 , S2 , . . . be arbitrary sets satisfying the assumptions of the theorem. Note that, by the first part of the proof, S is measurable. To prove (2.24), we consider two cases. Case 1: χS is integrable. Then χSn is integrable for every n ∈ N, because χSn ≤ χS , and we have χS  χS1 + χS2 + · · · . Then (2.24) can be proved as in (2.25). ∞   Case 2: µ(S) = ∞. Suppose that ∞ n=1 µ(Sn ) < ∞. Then ∞ n=1 χSn < ∞ and  thus there exists an integrable function f such that n=1 χSn = f a.e. But ∞ n=1 χSn = χS , and hence f = χS a.e.  This means that χS is integrable, contrary to the assumption. Therefore ∞
n=1 µ(Sn ) = ∞. The notion of the integral over an interval [a, b], introduced in Section 2.9, can be extended to the integral over any measurable set : 

 f=

f χ .



Integrability of the product f χ follows the form of the definition of measurable sets and Theorem 2.9.5. For any measurable set  ⊆ R, we can define the space L1 () of all integrable functions vanishing outside of . With the norm inherited from L1 (R), L1 () is a Banach space. In Section 2.4 the question whether integrability of |f | implies integrability of f is considered. The problem is closely related to the problem of existence of nonmeasurable sets, which in turn is related to existence of nonmeasurable functions. Using Theorem 2.11.7, one can prove the following: If f is a measurable function, then |f | ∈ L1 (R) implies f ∈ L1 (R).

Definition 2.11.5.

(Measurable functions) A function f is called measurable if there exists a sequence of step functions f1 , f2 , . . . such that fn → f a.e. Obviously, every integrable function is measurable. An easy argument shows that every locally integrable function is measurable. There are measurable functions which are not locally integrable. Indeed, consider the function defined by  1/x for x > 0, f (x) = 0 for x ≤ 0.

2.12 Complex-Valued Lebesgue Integrable Functions

71

Theorem 2.11.6. The measurable functions form a vector space. The absolute value of a measurable function is a measurable function. The product of measurable functions is a measurable function.

Proof: Let f and g be measurable functions, and let (fn ) and (gn ) be sequences of step functions such that fn → f a.e. and gn → g a.e. Then fn + gn → f + g a.e., λfn → λf a.e., |fn | → |f | a.e., fn gn → fg a.e. Since (fn + gn ), (λfn ), (fn gn ), and (|fn |) are sequences of step functions, the assertions follow.
Observe that Theorem 2.11.6 implies that if f and g are measurable functions, then max{f , g} and min{f , g} are measurable functions.

Theorem 2.11.7. If f is a measurable function and |f | ≤ g for some locally integrable function g, then f is locally integrable. Proof: Let [a, b) be an arbitrary bounded interval in R. We need to

show that f χ[a,b) is an integrable function. Suppose first that f ≥ 0. Let (fn ) be a sequence of step functions such that fn → f a.e. Define gn = χ[a,b) min{fn , g}. Then (gn ) is a sequence of integrable functions convergent to f χ[a,b) almost everywhere. Thus, by the Lebesgue dominated convergence theorem, f χ[a,b) is an integrable function. For an arbitrary measurable function f , we first define f + = 12 (|f | + f ) and f − = 12 (|f | − f ). Then we have f + , f − ≥ 0 and f = f + − f − . Therefore, the theorem follows from the first part of the proof and Theorem 2.11.6.
Other properties of the Lebesgue measure, measurable sets, and measurable functions can be found in the exercises.

2.12 Complex-Valued Lebesgue Integrable Functions In this section we extend the definition of Lebesgue integrable functions to include functions with complex values. First we define a complex-valued step function: f is a complex-valued step function if there exist complex numbers λ1 , . . . , λn and intervals [a1 , b1 ), . . . , [an , bn ) such that f (x) = λ1 χ[a1 ,b1 ) + · · · + λn χ[an ,bn ] .

(2.26)

72

Chapter 2

The integral



The Lebesgue Integral

f of the step function (2.26) is defined by  f = λ1 (b1 − a1 ) + · · · + λn (bn − an ).

The defined integral has the same properties as the integral of real step functions. Note that, if f is a step function, thenits real part  Re f and imaginary part Im f are step functions and we have f = Re f + i Im f .

Definition 2.12.1.

(Lebesgue integral for complex-valued functions) A complex-valued function f is Lebesgue integrable if there exists a sequence of step functions (fn ) such that the following two conditions are satisfied: (a)

∞  

|fn | < ∞;

n=1

(b)

f (x) =

∞ 

fn (x)

for every x ∈ R such that

n=1

∞    fn (x) < ∞. n=1

The integral of f is then defined by  f=

∞  

fn .

n=1

If a function f and a sequence of step functions (fn ) satisfy (a) and (b), then we write as before f

∞ 

fn

or

f  f1 + f2 + · · · .

n=1

We do not have to repeat the argument presented for real functions to prove that the defined integral does not depend on a particular representation by a series of step functions and thus is well defined. This follows immediately from the next theorem.

Theorem 2.12.2. A complex-valued function f is integrable if and only if its real part Re f and its imaginary part Im f are integrable. Moreover, if f is integrable, then    f = Re f + i Im f . (2.27)

2.12 Complex-Valued Lebesgue Integrable Functions

73

Proof: Let f  f1 + f2 + · · · , where f1 , f2 , . . . are complex-valued step functions. Then     |Re fn | ≤ |fn | and |Im fn | ≤ |fn |, and hence ∞  

|Re fn | < ∞

and

n=1

∞  

|Im fn | < ∞.

n=1

We can thus write Re f  Re f1 + Im f1 − Im f1 + Re f2 + Im f2 − Im f2 + · · · and Im f  Im f1 + Re f1 − Re f1 + Im f2 + Re f2 − Re f2 + · · · . This proves that Re f and Im f are integrable functions and (2.27) holds. If Im f  f1 + f2 + · · · and Re f  g1 + g2 + · · · , then f  f1 + ig1 + f2 + ig2 + · · · , which proves integrability of f .


Theorem 2.12.3. If a complex-valued  function  f is integrable, then the realvalued function |f | is integrable and | f | ≤ |f |.

Proof: Let f  f1 + f2 + · · · and sn = f1 + · · · + fn . Then,in view  of Corollary 2.7.11, sn → f a.e., and hence also |sn | → |f | a.e. Since ∞ n=1 |fn | < ∞, the ∞ function g = n=1 |fn | is integrable, and we have |sn | → |f | a.e.

and

|sn | < g.

Therefore, by the Lebesgue dominated convergence theorem, |f | is integrable. For every n ∈ N, we have            f1 + · · · + fn  =  sn  ≤ |sn |.       Hence, by letting n → ∞, we obtain | f | ≤ |f |.

Definition 2.12.4.




(Locally integrable complex-valued functions) A complex-valued function defined on R is called locally integrable if its real part and imaginary part are locally integrable. The next theorem follows immediately from Theorem 2.12.3.

74

Chapter 2

The Lebesgue Integral

Theorem 2.12.5. The absolute value |f | of a locally integrable complex-valued function f is locally integrable. At this point the reader should look back at the properties of real-valued integrable functions to see how they generalize to complex-valued functions. For example, the important Theorems 2.8.1, 2.8.2, and 2.8.4 remain true for complex-valued functions. Most of those results can be easily obtained from the real-valued cases by considering the real and imaginary parts of functions and using the theorems of this section.

2.13 The Spaces Lp (R) In this section, we introduce a family of spaces that are among the most important examples of Banach spaces.

(Lp (R)) For a real p > 1, by L (R) we denote the space of all complex-valued locally integrable functions f such that |f |p ∈ L1 (R).

Definition 2.13.1.

p

We will prove that, for every p > 1, Lp (R) with the norm defined by  f p =

1/p |f |

p

(2.28)

is a Banach space. First we need to show that Lp (R) is a vector space, then that (2.28) is a norm in Lp (R), and finally that the space is complete. Even the fact that Lp (R) is a vector space is not obvious. The approach we take is similar to the one presented in Chapter 1 and concerning the space lp (see Example 1.2.6).

Theorem 2.13.2. (Hölder’s inequality) Let 1 < p < ∞, 1 < q < ∞, and p1 + 1 q

= 1. If f ∈ Lp (R) and g ∈ Lq (R), then fg ∈ L1 (R) and fg 1 ≤ f p g q .

(2.29)

Proof: Since the inequality is obvious if f p = 0 or g q = 0, we assume that f p = 0 and g q = 0. It was established in Chapter 1 (see (1.1)) that for any non-negative numbers a and b, we have ab ≤

ap bq + . p q

2.13 The Spaces Lp (R)

75

For any x ∈ R, by taking a=

|f (x)| f p

and

b=

|g(x)| , g q

we obtain |f (x)g(x)| |f (x)|p |g(x)|q ≤ q . p + f p g q q g q p f p Integrating both sides of this inequality gives  |fg| 1 1 ≤ + = 1. f p g q p q


Theorem 2.13.3. (Minkowski’s inequality) Let 1 ≤ p < ∞. If f , g ∈ Lp (R),

then f + g ∈ Lp (R) and

f + g p ≤ f p + g p .

(2.30)

Proof: For any non-negative numbers a and b and p ≥ 1, we have   (a + b)p ≤ 2p−1 ap + bp . (We leave the proof of this elementary inequality as an exercise.) Hence,  p   |f + g|p ≤ |f | + |g| ≤ 2p−1 |f |p + |g|p , proving f + g ∈ Lp (R). We already know that (2.30) holds for p = 1, so we can assume that p > 1. Let q > 1 be such that p1 + 1q = 1. Since |f + g|p = |f + g||f + g|p−1 ≤ |f ||f + g|p−1 + |g||f + g|p−1

(2.31)

and |f + g|p−1 ∈ Lq (R), Hölder’s inequality implies |f ||f + g|p−1 , |g||f + g|p−1 ∈ L1 (R) and    |f ||f + g|p−1 ≤ f p |f + g|p−1 q  = f p

1/q |f + g|(p−1)q

p/q

= f p f + g p ,

since (p − 1)q = p. Similarly,  p/q |g||f + g|p−1 ≤ g p f + g p .

76

Chapter 2

The Lebesgue Integral

Now, from (2.31), we obtain   p p−1 f + g p ≤ |f ||f + g| + |g||f + g|p−1 p/q

p/q

≤ f p f + g p + g p f + g p   p/q = f p + g p f + g p . Since p − pq = 1, in equality (2.30) follows.




Now it is easy to check that Lp (R) is a vector space for every p ≥ 1 and that (2.28) is a norm in Lp (R). To be exact, we should say that (2.28) defines a norm in the space Lp (R) of equivalence classes of functions in Lp (R) that are equal almost everywhere, but we will not make such a distinction. (See the discussion in Section 2.6.)

Theorem 2.13.4. The space Lp (R) is complete for all 1 ≤ p < ∞. Proof: Let (fn ) be a Cauchy sequence in Lp (R) and let M > 0 be a fixed number. Then, by Hölder’s inequality, 

M −M

 |fm − fn | ≤

M

−M

= (2M)

1/p   |fm − fn |p

1/q

1/q

M −M

1

fm − fn p → 0 as m, n → ∞,

where p1 + 1q = 1. Thus, (fn ) is a Cauchy sequence in L1 ([−M, M]) and hence it converges to a function f ∈ L1 ([−M, M]), that is, 

M −M

|f − fn | → 0 as n → ∞.

By Theorem 2.8.2, the sequence (fn ) has a subsequence convergent to f almost everywhere in [−M, M]. Now we apply the described method for M = 1, 2, . . . . First we find a subsequence (f1,n ) of (fn ) that converges almost everywhere in [−1, 1]. Next, since (f1,n ) is a Cauchy sequence in Lp ([−2, 2]), it has a subsequence (f2,n ) that converges almost everywhere in [−2, 2]. Continuing in this fashion, we construct a collection of sequences (fm,n ), m = 1, 2, . . . , such that (fm,n ) is a subsequence of (fn ) for every m ∈ N, (fm+1,n ) is a subsequence of (fm,n ) for every m ∈ N, (fm,n ) converges almost everywhere in [−m, m], as n → ∞, for every m ∈ N.

2.13 The Spaces Lp (R)

77

Since (fn,n ) is a subsequence of (fn ), there exists a sequence of indices pn → ∞ such that fn,n = fpn for all n ∈ N. Clearly, (fpn ) is a subsequence of (fn ) that converges almost everywhere in R to some function f . Since (fpn ) is a Cauchy sequence in Lp (R), given any ε > 0, we have  |fpm − fpn |p < ε for sufficiently large m and n. By letting m → ∞, we obtain 

p

|f − fpn |p = f − fpn p ≤ ε, by Fatou’s lemma (Theorem 2.8.5). This proves that f ∈ Lp (R). Moreover, f − fn p ≤ f − fpn p + fpn − fn p < 2ε1/p for all sufficiently large n. Since ε is an arbitrary positive number, the sequence (fn ) converges to f in Lp (R).
If  is a measurable subset of R, then Lp () can be defined  as the space of all complex-valued measurable functions f on  such that  |f |p < ∞. The norm in Lp () is defined by  f p =

1/p |f |p

.



It is easy to see that the theorems in this section remain true if Lp (R) is replaced by Lp (). In fact, Lp () can be identified with the subspace of Lp (R) of all functions vanishing outside of . The space L∞ (R) is also used, but it is defined differently. First we introduce the essential supremum of a measurable function f , denoted by ess sup f , as the smallest number M such that f ≤ M almost everywhere. If no such number exists, then we write ess sup f = ∞. Then L∞ (R) is defined as the space of all measurable functions f such that ess sup |f | < ∞ and the norm in L∞ (R) is defined by f ∞ = ess sup |f |.

(2.32)

It is easy to see that Hölder’s and Minkowski’s inequalities hold when p = ∞ (and q = 1), and thus L∞ (R) is a vector space and (2.32) is a norm. It can also be shown that L∞ (R) is complete.

78

Chapter 2

The Lebesgue Integral

2.14 Lebesgue Integrable Functions on RN The theory presented in the first 13 sections of this chapter generalizes to functions defined on RN without any essential difficulties. Some changes are necessary in Section 2.2, but the remaining sections (with the exception of Section 2.10) are written in such a way that they can be read with RN in mind. We include in this section some basic definitions to indicate what changes are necessary in Section 2.2. By a semiopen interval in RN , we mean a set I, which can be represented as I = [a1 , b1 ) × · · · × [aN , bN ).

(2.33)

In other words, x = (x1 , . . . , xN ) ∈ I if ak ≤ xk < bk for k = 1, 2, . . . , N. For the interval (2.33), we define m(I) = (b1 − a1 ) · · · (bN − aN ). If the dimension N = 1 then m(I) is just the length of I; if N = 2 then m(I) is the area of I, if N = 3 then m(I) is the volume of I. By a step function we mean a finite linear combination of characteristic functions of semiopen intervals: f = λ1 χI1 + · · · + λn χIn .

(2.34)

For the step function f in (2.34), we define  f = λ1 m(I1 ) + · · · + λn m(In ).

(Lebesgue integrable function on RN ) A real or complex-valued function f defined on RN is called Lebesgue integrable if there exists a sequence of step functions (fn ) such that the following two conditions are satisfied: ∞   |fn | < ∞; (a)

Definition 2.14.1.

n=1

(b)

f (x) =

∞ 

fn (x)

for every x ∈ RN such that

n=1

∞ 

|fn (x)| < ∞.

n=1

The integral of f is then defined by  f=

∞  

fn .

n=1

The space of all Lebesgue integrable functions on RN is denoted by L1 (RN ).

2.14 Lebesgue Integrable Functions on RN

79

(Locally integrable function on RN ) A function f defined on RN is called locally integrable if for every bounded interval I , the product f χI is an integrable function.

Definition 2.14.2.

Lebesgue measure on RN can be defined as in Section 2.11: A set S ⊆ RN is called measurable if the characteristic function of S is a locally integrable function. If the characteristic function χS is an integrable function, then by the  measure of S, denoted by µ(S), we mean the value of the integral µ(S) = χS . If χS is locally integrable but not integrable, then we define µ(S) = ∞. A function f is called measurable if there exists a sequence of step functions f1 , f2 , . . . such that fn → f a.e.  For a measurable set  ⊆ RN , by  f we mean  theintegral of the function equal to f on  and 0 everywhere else, that is,  f = f χ .

(Lp (RN )) For a real p > 1, by L (RN ), we denote the space of all complex-valued locally integrable functions f on RN such that |f |p ∈ L1 (RN ).

Definition 2.14.3.

p

For every p > 1, Lp (RN ) with the norm defined by  f p =

1/p |f |p

(2.35)

is a Banach space and Hölder’s inequality holds: If 1 < p < ∞, 1 < q < ∞, 1 1 p N q N 1 N p + q = 1, f ∈ L (R ), and g ∈ L (R ), then fg ∈ L (R ) and fg 1 ≤ f p g q .

(2.36)

The space of all functions f ∈ Lp (RN ) vanishing outside a measurable set  ⊆ RN is denoted by Lp (). It is a vector subspace of Lp (RN ). If  is of finite measure, then Lp () ⊂ Lr () whenever p ≥ r. The essential supremum of a measurable function f , denoted by ess sup f , is the smallest number M such that f ≤ M almost everywhere. If no such number exists, then we write ess sup f = ∞. Then L∞ (RN ) is defined as the space of all measurable functions f such that ess sup |f | < ∞ and the norm in L∞ (RN ) is defined by f ∞ = ess sup |f |. Applications frequently involve multiple integrals. Changing the order of integration is crucial in many arguments. For example, given (a, b) and (c, d),

80

Chapter 2

The Lebesgue Integral

intervals in R, and f , a function on (a, b) × (c, d), we want to be able to use the equality  d b  b d f (x, y)dydx = f (x, y)dxdy. a

c

c

a

For the Riemann integral, the equality is usually proved under the assumption of continuity of f , or at least piecewise continuity. The Lebesgue integral allows an essential relaxation of these assumptions. In the following theorem, we show that integrability of f over (a, b) × (c, d) is sufficient. The theorem is a special case of Fubini’s theorem (Guido Fubini (1879–1943)). The formulation of the theorem requires clarification of the notation. Let (a, b) and (c, d) be bounded or unbounded intervals in R and let J =  (a, b) × (c, d). By the integral J f , we mean the integral of f as defined in this section (Definition 2.14.1). For a fixed x ∈ (a, b) the integral 

d

f (x, y)dy

(2.37)

c

denotes the Lebesgue integral of the function fx (y) = f (x, y), that is, d d c f (x, y)dy = c fx . If the function fx is integrable over (c, d) for every x ∈ (a, b), then (2.37) defines a function on (a, b): 

d

F(x) =

f (x, y)dy. c

If F ∈ L1 ((a, b)), then F can be integrated over (a, b). The resulting integral will be denoted as a double integral of f : 

b

F=

 b

d

f (x, y)dydx.

a

a

c

Notice that the question whether F is integrable or not makes sense even if F is defined almost everywhere, that is, if the integral (2.37) exists almost everywhere in (a, b). In a similar fashion, only changing the order, we define the integral  d

b

f (x, y)dxdy. c

a

Theorem 2.14.4. (Fubini’s theorem) Let (a, b) and (c, d) be bounded or unbounded intervals in R and let J = (a, b) × (c, d). If f is an integrable function on J, then the function  d F(x) = f (x, y)dy c

2.14 Lebesgue Integrable Functions on RN

81

is defined almost everywhere in (a, b), F is integrable in (a, b), and we have 



d

f= J

F=

 b

c

d

f (x, y)dydx. a

(2.38)

c

Proof: Since f is integrable over J, there are intervals In = [an , bn ) × [cn , dn ) ⊆ J and scalars λ1 , λ2 , . . . such that f  λ1 χI1 + λ2 χI2 + · · · . If we denote gn = χ[an ,bn ) ,

hn = χ[cn ,dn ) ,

and

fn (x, y) = gn (x)hn (y),

then we have f  λ1 f 1 + λ2 f 2 + · · · . Note that



 fn =

(2.39)

 gn

hn .

 (Remember that fn ∈ L1 (R2 ) and gn , hn ∈ L1 (R), so that the symbol has different meanings on the left-hand side and the right-hand side of the above equality.) Since       |λ1 | g1 h1 + |λ2 | g2 h2 + · · · = |λ1 | f1 + |λ2 | f2 + · · · < ∞, there is an integrable function F (a function of x) such that       F  λ1 h1 g1 + λ2 h2 g2 + · · · .

(2.40)

Let x0 ∈ (a, b) be a point where the series (2.40) converges absolutely. Then there exists an integrable function G (a function of y) such that     G  λ1 g1 (x0 ) h1 + λ2 g2 (x0 ) h2 + · · · . (2.41) If the series (2.41) converges absolutely at some y ∈ (c, d), then λ1 g1 (x0 )h1 (y) + λ2 g2 (x0 )h2 (y) + · · · = f (x0 , y), by (2.39). Thus, the function f (x0 , y) is integrable over (c, d), and we have 

d c

 f (x0 , y)dy = λ1 g1 (x0 )

 h1 + λ2 g2 (x0 )

h2 + · · ·

(2.42)

82

Chapter 2

The Lebesgue Integral

for every x0 , where the series (2.40) converges absolutely. Consequently, the function  d F(x) = f (x, y)dy c

is defined almost everywhere in (a, b), and the first part of the theorem is proved. Now, in view of (2.40), F is integrable and we have 

b

F=

a

 b a

d

 f (x, y)dydx = λ1

 g1

 h1 + λ2

 g2

h2 + · · · .

c

On the other hand,      λ1 g1 h1 + λ2 g2 h2 + · · · = f . J

Thus, (2.38) is proved.




2.15 Convolution  If the integral f (x − y)g(y)dy exists for all x ∈ R, or at least almost everywhere, then it defines a function which is called the convolution of f and g, denoted by f ∗ g. In this section, we discuss some basic properties of convolution.

Theorem 2.15.1. If f , g ∈ L1 (R), then the function f (x − y)g(y) is integrable for

almost all x ∈ R. Moreover, the convolution  (f ∗ g)(x) = f (x − y)g(y)dy is an integrable function, and we have    |f ∗ g| ≤ |f | |g|.

(2.43)

Proof: First we show that the function H(x, y) = f (x − y)g(y) is integrable

in R2 . Since f , g ∈ L1 (R), there are step functions f1 , f2 , . . . and g1 , g2 , . . . such that f  f1 + f2 + · · ·

and

g  g1 + g1 + · · · .

Now define Fn (x, y) = fn (x − y)

and

Gn (x, y) = gn (y).

2.15 Convolution

83

We will show that H

∞  n 

Fk Gn−k+1 .

(2.44)

n=1 k=1

In fact, since      n n  n       ≤ F G G | ≤ | |gn−k+1 | |F |f  k n−k+1  k n−k+1 k   k=1

and the series

k=1

∞  

(2.45)

k=1

|fn |

and

n=1

∞  

|gn |

n=1

are absolutely convergent, we have  n  ∞       Fk Gn−k+1  < ∞.    n=1

k=1

Moreover, if for some (x, y) ∈ RN the series ∞  n 

Fk (x, y)Gn−k+1 (x, y)

n=1 k=1

converges absolutely and, for some n0 , m0 ∈ N, we have fn0 (x − y) = 0 and gm0 (y) = 0, then both series ∞ 

fn (x − y)

n=1

and

∞ 

gn (y)

n=1

converge absolutely. If, on the other hand, fn (x − y) = 0 for all n ∈ N or gn (y) = 0 for all n ∈ N, then H(x, y) = 0. In either case we have H(x, y) =

n ∞  

Fk (x, y)Gn−k+1 (x, y),

n=1 k=1 1 2 at that point (x, y). Therefore, (2.44) holds,  which means that H ∈ L (R ). By Fubini’s Theorem 2.14.4, the integral f (x − y)g(y)dy is defined almost everywhere in R (with respect to x) and defines an integrable function, that is, f ∗ g ∈ L1 (R). To show that (2.43) holds, note that since H is integrable, so is |H| and thus

84

Chapter 2



The Lebesgue Integral       |f ∗ g| =  f (x − y)g(y)dydx     f (x − y)g(y)dydx ≤

 =  =

   f (x − y)g(y)dxdy

  f (x − y)dx



(by Fubini’s theorem)

  g(y)dy



 |f (x − y)|dx is independent of y

 =

 |f |

|g|.




Theorem 2.15.2. If f , g ∈ L1 (R), then f ∗ g = g ∗ f . Proof follows easily by the change of variables (Theorem 2.10.4).

Theorem 2.15.3. If f is an integrable function and g is a bounded, locally integrable function, then the convolution f ∗ g is a continuous function. Proof: First observe that since |f (x − y)g(y)| ≤ M|f (x − y)| for some con stant M and every x ∈ R the integral f (x − y)g(y)dy is defined at every x ∈ R, by Theorem 2.9.5. Now we will show that the convolution f ∗ g is a continuous function. For any x, t ∈ R, we have        (f ∗ g)(x + t) − (f ∗ g)(x) =  f (x + t − y)g(y)dy − f (x − y)g(y)dy         =  f (x + t − y) − f (x − y) g(y)dy     ≤ f (x + t − y) − f (x − y)g(y)dy  ≤M

  f (t − y) − f (−y)dy.

 Now Theorem 2.4.3 implies limt→∞ |f (t − y) − f (−y)|dy = 0, proving continuity of f ∗ g.


2.16 Exercises 1. Denote by A the family consisting of all finite unions of semiopen intervals [a, b) and the empty set. Prove the following properties of A:

2.16 Exercises

85

(a) If A1 , . . . , An ∈ A, then A1 ∪ · · · ∪ An ∈ A. (b) If A1 , . . . , An ∈ A, then A1 ∩ · · · ∩ An ∈ A. (c) If A, B ∈ A, then A \ B ∈ A. 2. Show that step functions form a vector space. 3. Prove that the integral of a step function is independent of a particular representation (2.1). 4. Show that for any step functions f and g we have the following: (a) supp(f + g) ⊆ supp f ∪ supp g. (b) supp fg = supp f ∩ supp g. (c) supp |f | = supp f . (d) supp λf = supp f , λ ∈ R, λ = 0. (e) If |f | ≤ |g|, then supp f ⊆ supp g. 5. Prove Lemma 2.2.3. 6. Complete the argument in Example 2.2.5. 7. Show that, if f is integrable,   then the translated function τz f is integrable for any z ∈ R and τz f = f . 8. Prove Theorem 2.4.3 for step functions. 9. Expand the following functions into a series of step functions, that is, find step functions f1 , f2 , . . . such that f  f1 + f2 + · · · :  1 if x = 0, (a) f (x) = 0 if x = 0.  1 if x ∈ [a, b], (b) f (x) = a < b. 0 if x ∈ / [a, b], (c) f (x) = max{0, 1 − |x|}. (d) f is a piecewise continuous function with bounded support. 10. Show that f ∈ L1 (R) if and only if there are intervals [a1 , b1 ), [a2 , b2 ), . . . and numbers λ1 , λ2 , . . . such that f  λ1 χ[a1 ,b1 ) + λ2 χ[a2 ,b2 ) + · · · . 11. Show that if f  f1 + f2 + · · · then f + g  g + f1 + f2 + · · · for any step function g. In particular, if f  f1 + f2 + · · · then f − f1 − · · · − fn  fn+1 + fn+2 + · · · .

86

Chapter 2

The Lebesgue Integral

12. If f ∈ L1 (R) and f vanishes outside of a bounded interval [a, b), then there are step functions f1 , f2 , . . . vanishing outside of [a, b) such that f  f1 + f 2 + · · · . 13. If f ∈ L1 (R) and f ≥ 0, is it always possible to find non-negative step functions f1 , f2 , . . . such that f  f1 + f2 + · · · ? 14. Show that the characteristic function of the set of all rational numbers is Lebesgue integrable but not Riemann integrable. 15. Define f + = max{0, f } and f − = max{0, −f }. Prove that f ∈ L1 (R) if and only if f + ∈ L1 (R) and f − ∈ L1 (R). 16. Show that, if f is a continuous, integrable function, then there are step functions f1 , f2 , . . . such that f  f1 + f2 + · · · and |f |  |f1 | + |f2 | + · · · . 17. By a tent function we mean a function of the form    2(x − a)/(b − a) f (x) = 2(b − x)/(b − a)   0

if a ≤ x ≤ (a + b)/2, if (a + b)/2 ≤ x ≤ b, otherwise,

where a < b (see Figure 2.3). Show that f ∈ L1 (R) if and only if there exist tent functions f1 , f2 , . . . and numbers λ1 , λ2 , . . . such that f  λ1 f1 + λ2 f 2 + · · · .  18. Show that the relation f ∼ g if |f − g| = 0 is an equivalence in L1 (R). 19. Prove the following: (a) Every countable subset of R is a null set. (b) A countable union of null sets is a null set. 20. Consider the following sequence of subsets of the interval [0, 1]: S0 = [0, 1], ! 1 2  , , S1 = [0, 1] 3 3

Figure 2.3 A tent function.

2.16 Exercises       ! 1 2 4 5 7 8 , ∪ , ∪ , , S2 = [0, 1] 9 9 9 9 9 9

87

.. . ! 3n−1 −1  1 + 3k 2 + 3k  Sn = [0, 1] , , 3n 3n k=0

.. . Define C=

∞ "

Sn .

n=0

This set is called the Cantor set (Georg Ferdinand Ludwig Philipp Cantor (1845–1918)). (a) Show that C is a null set. (b) Show that C is not countable. 21. Prove the following properties of convergence almost everywhere: (a) If fn → f a.e. and λ ∈ R, then λfn → λf a.e. (b) If fn → f a.e. and gn → g a.e., then fn + gn → f + g a.e. (c) If fn → f a.e., then |fn | → |f | a.e. 22. Show that every Lebesgue integrable function can be approximated in norm and almost everywhere by a sequence of continuous functions. 23. Let f ∈ L1 (R). Define  fn (x) =

f (x) 0

if |x| ≤ n, otherwise.

Show that fn → f i.n. 24. Show that there exists an unbounded continuous function f ∈ L1 (R). 25. Show that if f is a uniformly continuous function on R and f ∈ L1 (R), then f is bounded and lim|x|→∞ f (x) = 0. 26. Show that locally integrable functions form a vector space. 27. Let f ∈ L1 (R) and let g be a bounded, locally  integrable function. Show 1 that fg ∈ L (R) and |fg| ≤ supx∈R |g(x)| |f |. 28. Show that the space L1 (J) is complete for any interval J ⊆ R.

88

Chapter 2

The Lebesgue Integral

29. Prove: If a sequence of locally integrable functions (fn ) converges almost everywhere to a function f and |fn | ≤ h for every n ∈ N, where h is a locally integrable function, then f is locally integrable. 30. In Example 2.7.9 we define a sequence of functions (fn ) convergent to 0 in norm but divergent at every point of [0, 1]. Find a subsequence of (fn ) convergent to 0 almost everywhere. 31. Prove: If a sequence of integrable functions (fn ) converges almost everywhere to a function f and |fn (x)| ≤ h(x) for almost all x ∈ R, all n ∈ N, and some integrable function h, then f is integrable and fn → f i.n. 32. Show that the function

 sin x f (x) =

1

if x = 0,

x

if x = 0

is  ∞not a Lebesgue integrable, although the improper Riemann integral −∞ f (x)dx converges. 33. Let M denote the collection of all measurable subsets of R. Prove the following: (a) ∅, R ∈ M. (b) If A1 , A2 , . . . ∈ M, then (c) If A1 , A2 , . . . ∈ M, then

∞

n=1 An

∈ M.

n=1 An

∈ M.

#∞

(d) If A, B ∈ M, then A\B ∈ M. (e) Intervals are measurable sets. (f) Open subsets of R are measurable. (g) Closed subsets of R are measurable. 34. Let M be the collection of all measurable subsets of R and let µ be the Lebesgue measure on R. Prove the following: (a) If A1 , A2 , . . . ∈ M, then  µ

∞  n=1

 An ≤

∞ 

µ(An ).

n=1

(b) If A, B ∈ M and A ⊆ B, then µ(B \ A) = µ(B) − µ(A). (c) If A1 , A2 , . . . ∈ M and A1 ⊆ A2 ⊆ A3 ⊆ · · · , then  ∞   µ An = lim µ(An ). n=1

n→∞

2.16 Exercises

89

(d) If A1 , A2 , . . . ∈ M and A1 ⊇ A2 ⊇ A3 ⊇ · · · , then  µ

∞ "

 An = lim µ(An ).

n=1

n→∞

35. Let f be a real-valued function on R. Show that the following conditions are equivalent: (a) f is measurable. (b) {x ∈ R: f (x) ≤ α} is a measurable set for all α ∈ R. (c) {x ∈ R: f (x) < α} is a measurable set for all α ∈ R. (d) {x ∈ R: f (x) ≥ α} is a measurable set for all α ∈ R. (e) {x ∈ R: f (x) > α} is a measurable set for all α ∈ R. 36. Prove: Let A1 , A2 , . . . be measurable sets such  that limn→∞ µ(An ) = 0. Then, for every f ∈ L1 (R), we have limn→∞ An f = 0. 37. Prove that L1 () is a Banach space forany measurable set  ⊆ RN . (The norm in L1 () is defined by f =  |f | and functions equal almost everywhere are identified.) 38. Show that for any non-negative numbers a and b and p ≥ 1, we have (a + b)p ≤ 2p−1 (ap + bp ). 39. Let

 g(x) =

√ 1/ x 0

for 0 < |x| < 1, otherwise.

/ L1 (R). Show that g ∈ L1 (R) but g 2 ∈ 40. Let f (x) = min{1, 1/|x|}. Show that f ∈ / L1 (R) but f ∈ L2 (R). 41. Show that Lp ([a, b]) ⊆ Lr ([a, b]) for any 1 ≤ r < p ≤ ∞. 42. Show that Hölder’s and Minkowski’s inequalities hold when p = ∞ and q = 1. 43. Let fn , f ∈ L∞ (R). Show that fn − f ∞ → 0 if and only if there exists a set of measure zero S ⊂ R such that fn → f uniformly on R\S. 44. Let f , g, h ∈ L1 (R). Prove the following properties of the convolution: (a) (f ∗ g) ∗ h = f ∗ (g ∗ h) (Associative), (b) (αf + βg) ∗ h = α(f ∗ h) + β(g ∗ h) (Distributive) where α and β are constants.

90

Chapter 2

The Lebesgue Integral

45. Calculate (f ∗ g)(x) for the following functions: (a) f (x) = sin x, g(x) = e−|x| . (b) f (x) = xχ[−2,1] (x), g(x) = x2 + 2. (c) f (x) = ex , g(x) = χ[0,∞] (x). (d) f (x) = χ[−4,0] (x), g(x) = χ[−1,1] (x). 46. (a) If (f , g) ∈ L1 (R), show that (f ∗ g)(x) exists for almost all x and (f ∗ g) ∈ L1 (R). ∞ ∞ ∞ (b) Prove that −∞ (f ∗ g)(x)dx = −∞ f (t)dt −∞ g(s)ds and f ∗ g 1 ≤ f 1 g 1 . 1 47. If g(x) = 2a H(a − |x|), show that (f ∗ g)(x) is the average value of f in [x − a, x + a].

48. Let f be the characteristic function of the interval [−1, 1]. Calculate the convolutions f ∗ f and f ∗ f ∗ f . 49. Let f ∈ L1 (R) and let g be a bounded differentiable function on R. Show that f ∗ g is differentiable. If, in addition, g is bounded show that (f ∗ g) = f ∗ g . 50. Let f be a locally integrable function on R and let g be a differentiable function with bounded support in R. Show that f ∗ g is differentiable and (f ∗ g) = f ∗ g . 51. Prove the following properties of differentiability of the convolution: (a) (f ∗ g)

= f ∗ g . (b) (f ∗ g)(n+m) (x) = f (n) (x) ∗ g (m) (x). 52. (Young’s inequality) Suppose 1 ≤ p, q, r ≤ ∞ satisfy the equation 1 1 1 + =1+ . p q r If f ∈ Lp (R) and g ∈ Lq (R), show that f ∗ g ∈ Lr (R) and we have f ∗ g r ≤ f p g q (William Henry Young (1863–1942)). 53. Prove that L∞ (RN ) is complete. 54. Show that Lp (RN ) is separable for 1 ≤ p < ∞. 55. Show that L∞ ([0, 1]) is not separable.

2.16 Exercises

91

56. Show that the space Cc (RN ) of continuous functions with compact support is dense in Lp (RN ) for 1 ≤ p < ∞. 57. Show that C ([0, 1]) is not dense in L∞ ([0, 1]). 58. Show that the space D (RN ) of smooth functions with compact support is dense in Lp (RN ) for 1 ≤ p < ∞.

This page intentionally left blank

Hilbert Spaces and Orthonormal Systems

CHAPTER

3

“As long as a branch of knowledge offers an abundance of problems, it is full of vitality.” David Hilbert “Hilbert spaces constitute at present the most important examples of Banach spaces, not only because they are the most natural and closest generalization in the realm of ‘infinite dimensions,’ of our classical Euclidean geometry, but chiefly for the fact they have been, up to now, the most useful spaces in the applications to functional analysis.” Jean Dieudonné

3.1 Introduction The theory of Hilbert spaces was initiated by David Hilbert (1862–1943) in his 1912 work on quadratic forms in infinitely many variables, which he applied to the theory of integral equations. He published a series of papers on the subject during 1904–1910 and formulated the theory of the space l2 (square summable sequences) in connection with integral equations. The idea of the l2 -space and the spectral theory of bounded quadratic forms were published in 1906. All these papers were reproduced in the form of a book entitled Grundzüge einer allgemeinen Theorie der linearen Integralgleichungen in 1912. The book had a tremendous influence on mathematical analysis and its applications. Years later, John von Neumann (1903–1957) first formulated an axiomatic theory of Hilbert spaces and developed the modern theory of operators on Hilbert spaces. His remarkable contribution to this area has provided the mathematical foundation of quantum mechanics. Von Neumann’s work has

93

94

Chapter 3

Hilbert Spaces and Orthonormal Systems

also provided a physical interpretation of quantum mechanics in terms of abstract relations in an infinite dimensional Hilbert space. This chapter is concerned with inner product spaces (called also pre-Hilbert spaces or unitary spaces) and Hilbert spaces. The basic ideas and properties will be discussed with special attention given to orthonormal systems. The theory is illustrated by numerous examples.

3.2 Inner Product Spaces Definition 3.2.1.

(Inner product space) Let E be a complex vector space. A mapping ·, · : E × E → C is called an inner product in E if for any x, y, z ∈ E and α, β ∈ C the following conditions are satisfied: (a) x, y = y, x (the bar denotes the complex conjugate); (b) αx + βy, z = αx, z + βy, z; (c) x, x ≥ 0; (d) x, x = 0 implies x = 0. A vector space with an inner product is called an inner product space. According to the definition, the inner product of two vectors is a complex number. By (a), x, x = x, x, which means that x, x is a real number for every x ∈ E. It follows from (b) that x, αy + βz = αy + βz, x = αy, x + βz, x = αx, y + βx, z. In particular αx, y = αx, y

and

x, αy = αx, y.

Hence, if α = 0, we have 0, y = x, 0 = 0.

Example 3.2.2. The simplest, although important, example of an inner product space is the space of complex numbers C. The inner product is defined by x, y = xy.
Example 3.2.3. The space CN of ordered N-tuples x = (x1 , . . . , xN ) of complex numbers, with the inner product defined by x, y =

N  k=1

xk y k ,

x = (x1 , . . . , xN ), y = (y1 , . . . , yN ),

3.2 Inner Product Spaces

is an inner product space.

95


Example 3.2.4.The space l2 of all sequences (x1 , x2 , x3 , . . .) of complex num-

bers such that fined by

x, y =

∞ 2 k=1 |xk |

∞ 

xk y k ,

< ∞ (see Section 1.2), with the inner product de-

x = (x1 , x2 , x3 , . . .), y = (y1 , y2 , y3 , . . .),

k=1

is an infinite dimensional inner product space. As we will see later, this space is, in a sense, the most important example of an inner product space (see Theorem 3.4.27).


Example 3.2.5. Consider the space of sequences (x1 , x2 , x3 , . . .) of complex numbers with only a finite number of nonzero terms. This is an inner product space with the inner product defined as in Example 3.2.4.
Example 3.2.6. The space C ([a, b]) of all continuous complex valued functions on the interval [a, b], with the inner product  f , g =

b

f (x)g(x)dx, a

is an inner product space.




Example 3.2.7. The space L2 (R) (see Section 2.13) with the inner product defined by



∞

−∞

f (x)g(x)dx,

and, more generally, the space L2 (RN ) (see Section 2.14) with the inner product defined by  f (x)g(x)dx, RN

are very important inner product spaces. In applications we often use a subset  of RN and the space L2 () with the inner product defined by the integral over . For example, the space L2 ([a, b]) is the right setting in many cases.


Example 3.2.8. Let E be the Cartesian product of inner product spaces E1 and E2 , that is, E = E1 × E2 = {(x, y): x ∈ E1 , y ∈ E2 }. The space E is an inner product space with the inner product defined by % $ (x1 , y1 ), (x2 , y2 ) = x1 , x2  + y1 , y2 .

96

Chapter 3

Hilbert Spaces and Orthonormal Systems

Note that E1 and E2 can be identified with the subspaces E1 × {0} and {0} × E2 , respectively. Similarly, we can define the inner product on E1 × · · · × En . This method can be used to construct new examples of inner product spaces.
An inner product space is a vector space with an inner product. It turns out that every inner product space is also a normed space with the norm defined by x = x, x. First notice that this functional is well defined because x, x is always a nonnegative (real) number and x = 0 if and only if x = 0. Moreover,

λx = λx, λx = λλx, x = |λ| x . It thus remains to prove the triangle inequality. This is not as simple as the first two conditions. We first prove the so-called Schwarz’s inequality (Hermann Amandus Schwarz (1843–1921)), which will be used in the proof of the triangle inequality.

Theorem 3.2.9. (Schwarz’s inequality) For any two elements x and y of an inner product space, we have

  x, y ≤ x y .

(3.1)

The equality |x, y| = x y holds if and only if x and y are linearly dependent.

Proof: If y = 0, then (3.1) is satisfied because both sides are equal to zero. Assume then y = 0. We have

0 ≤ x + αy, x + αy = x, x + αx, y + αy, x + |α|2 y, y.

(3.2)

Now put α = −x, y/y, y in (3.2), and then multiply by y, y to obtain  2 0 ≤ x, xy, y − x, y . This gives Schwarz’s inequality. If x and y are linearly dependent, then x = αy for some α ∈ C. Hence,     x, y = x, αx = |α|x, x = |α| x x = x αx = x y . Now, let x and y be vectors such that |x, y| = x y , or, equivalently, x, yy, x = x, xy, y.

(3.3)

We will show that y, yx−x, yy = 0, which proves that x and y are linearly dependent. Indeed, by (3.3) we have

3.2 Inner Product Spaces

$

y, yx − x, yy, y, yx − x, yy

97

%

= y, y2 x, x − y, yy, xx, y − x, yy, yy, x + x, yy, xy, y = 0, completing the proof.




Corollary 3.2.10. (Triangle inequality) For any two elements x and y of an inner product space we have x + y ≤ x + y .

(3.4)

Proof: When α = 1, Equation (3.2) can be written as x + y 2 = x + y, x + y = x, x + 2 Rex, y + y, y ≤ x, x + 2|x, y| + y, y ≤ x 2 + 2 x y + y 2  2 = x + y ,

(by Schwarz’s inequality)

where Re z denotes the real part of z ∈ C.




The preceding discussion justifies the following definition:

Definition 3.2.11.

(Norm in an inner product space) By the norm in an inner product space E we mean the functional defined by x = √ x, x. We have proved that every inner product space is a normed space. It is only natural to ask whether every normed space is an inner product space. More precisely: is it possible√to define in a normed space (E, · ) an inner product ·, · such that x = x, x for every x ∈ E? In general, the answer is negative. In the following theorem, we prove a property of the norm in an inner product space, which is a necessary and sufficient condition for a normed space to be an inner product space; see Exercise 12.

Theorem 3.2.12. (Parallelogram law) For any two elements x and y of an inner product space, we have   x + y 2 + x − y 2 = 2 x 2 + y 2 . (3.5) Proof: We have x + y 2 = x + y, x + y = x, x + x, y + y, x + y, y

98

Chapter 3

Hilbert Spaces and Orthonormal Systems

and hence x + y 2 = x 2 + x, y + y, x + y 2 .

(3.6)

Now replace y by −y in (3.6) to obtain x − y 2 = x 2 − x, y − y, x + y 2 . By adding (3.6) and (3.7), we obtain the parallelogram law.

(3.7)


One of the most important consequences of having the inner product is the possibility of defining orthogonality of vectors. This makes the theory of Hilbert spaces very different from the general theory of Banach spaces.

Definition 3.2.13.

(Orthogonal vectors) Two vectors x and y in an inner product space are called orthogonal, denoted by x ⊥ y, if x, y = 0. If x ⊥ y, then y, x = x, y = 0, and thus y ⊥ x. In other words, the relation ⊥ is symmetric. The next theorem is another example of the geometric character of the norm defined by an inner product.

Theorem 3.2.14. (Pythagorean formula) For any pair of orthogonal vectors, we have x + y 2 = x 2 + y 2 .

(3.8)

Proof: If x ⊥ y, then x, y = y, x = 0, and thus the equality follows immediately from (3.6).
In the definition of the inner product space we assume that E is a complex vector space. It is possible to define a real inner product space with the inner product of any two vectors being a real number. Then condition (b) in Definition 3.2.1 becomes x, y = y, x. All the preceding theorems hold in the real case. If, in Examples 3.2.2–3.2.7, the word complex is replaced by real and C by R, we obtain a number of examples of real inner product spaces. A finite dimensional real inner product space is called a Euclidean space. N If x = (x1 , . . . , xN ) and N y = (y1 , . . . , yN ) are vectors in R , then the inner product x, y = k=1 xk yk can be defined equivalently as x, y = x y cos θ , where θ is the angle between vectors x and y. In this case, Schwarz’s inequality follows from |x, y| = | cos θ | ≤ 1, x y

x = 0, y = 0.

3.3 Hilbert Spaces

99

Figure 3.1 A “typical” function fn in Example 3.3.5.

3.3 Hilbert Spaces Definition 3.3.1. (Hilbert space) A complete inner product space is called a Hilbert space. By completeness of an inner product space E, we mean completeness of E as a normed space with the norm defined by the inner product. Now we are going to discuss completeness of the inner product spaces mentioned in Section 3.2 and also give some new examples of inner product spaces and Hilbert spaces.

Example 3.3.2. Since C is complete, it is a Hilbert space, and so is CN .




Example 3.3.3. l2 is a Hilbert space. The completeness was proved in Section 1.4 (see Example 1.4.6).




Example 3.3.4. The space E described in Example 3.2.5 is an inner product space, which is not a Hilbert space. It is not complete. The sequence   1 1 1 xn = 1, , , . . . , , 0, 0, . . . 2 3 n is a Cauchy sequence: & lim xn − xm = lim

n,m→∞

n,m→∞

max{m,n} 

k=min{m,n}+1

1 k2

'1/2 = 0.

However, the sequence does not converge in E, because its limit (1, 12 , 13 , . . .) is not in E. (The sequence (xn ) converges in l2 .)


Example 3.3.5. The space discussed in Example 3.2.6 is another example of an incomplete inner product space. In fact, consider the following sequence of

100

Chapter 3

Hilbert Spaces and Orthonormal Systems

functions in C ([0, 1]), (see Figure 3.1):  if 0 ≤ x ≤ 12 ,  1   1 + 12 , if 12 ≤ x ≤ 2n fn (x) = 1 − 2n x − 12   1 + 12 ≤ x ≤ 1. 0 if 2n Evidently, fn ’s are continuous. Moreover, 

1 1 + fn − fm ≤ n m

1/2 → 0,

as m, n → ∞.

Thus, (fn ) is a Cauchy sequence. It is easy to see that the sequence is pointwise convergent to the function  1 if 0 ≤ x ≤ 12 , f (x) = 0 if 12 < x ≤ 1. The limit function is not continuous and thus not an element of C ([0, 1]). Therefore the sequence (fn ) is not convergent in C ([0, 1]). Consequently, C ([0, 1]) is not a Hilbert space.


Example 3.3.6. The spaces L2 (R) and L2 ([a, b]) are Hilbert spaces (see Theorem 2.13.4).




Example 3.3.7. Let ρ be a measurable function defined on the interval [a, b] such that ρ(x) > 0 almost everywhere in [a, b]. Denote by L2,ρ ([a, b]) the space of all complex-valued measurable functions on [a, b] such that 

b

 f (x)2 ρ(x)dx < ∞.

a

This is a Hilbert space with the inner product 

b

f , g =

f (x)g(x)ρ(x)dx. a

To prove completeness, consider a Cauchy sequence (fn ) in L2,ρ ([a, b]). Then  fm − fn 2L2,ρ ([a,b]) = Define

b

 fm (x) − fn (x)2 ρ(x)dx → 0

a

√ Fn = fn ρ,

n ∈ N.

as n → ∞.

3.3 Hilbert Spaces

101

Since  Fm − Fn 2L2 ([a,b])

b

 Fm (x) − Fn (x)2 dx

= 

a

 fm (x) ρ(x) − fn (x) ρ(x) 2 dx

b

= 

a b

 fm (x) − fm (x)2 ρ(x)dx

= a

= fm − fn 2L2,ρ ([a,b]) , (Fn ) is a Cauchy sequence in L2 ([a, b]). Thus, there exists F ∈ L2 ([a, b]) such that  b   2 Fn (x) − F(x)2 dx → 0. Fn − F L2 ([a,b]) = a

We can show that ∈L 2,ρ pleteness of L ([a, b]). √F ρ

2,ρ

([a, b]) and fn →

√F ρ

in L2,ρ ([a, b]), proving com


Example 3.3.8. (Sobolev spaces) Let  be an open set in RN . Denote by

˜ m (), m = 1, 2, . . . , the space of all complex-valued functions f ∈ C m () H such that Dα f ∈ L2 () for all |α| ≤ m, where α = (α1 , . . . , αN ), α1 , . . . , αN are non-negative integers, |α| = α1 + · · · + αN , and Dα f =

∂ |α| f αN . ∂x1α1 ∂x2α2 . . . ∂xN

For example, if N = 2, α = (2, 1), we have Dα f =

∂ 3f . ∂x12 ∂x2

˜ m (), we thus have For f ∈ H 2   |α|    α ∂α f   ∂x 1 ∂x 2 . . . ∂xαN  < ∞ 

1

2

N

for every multi-index α = (α1 , α2 , . . . , αN ) such that |α| ≤ m. The inner prod˜ m () is defined by uct in H   Dα f Dα g. f , g =  |α|≤m

102

Chapter 3

Hilbert Spaces and Orthonormal Systems

˜ 2 () is If  ⊆ R2 , then the inner product in H  ( f g + fx gx + fy gy + fxx gxx + fyy gyy + fxy gxy ).

 f , g = 

˜ m (a, b) is If  = (a, b) ⊂ R, the inner product in H  b m dn f dn g  f , g = . n n a n=0 dx dx ˜ m () is an inner product space, but it is not a Hilbert space because it H ˜ m (), denoted by H m (), is a Hilbert is not complete. The completion of H m space. The space H () is a particular case of a general class of spaces denoted by W m,p () (see Section 6.3), introduced by Sergei Lvovich Sobolev (1908– 1989). We have H m () = W m,2 (). Because of the applications to partial differential equations, spaces H m () belong to the most important examples of Hilbert spaces.
Since every inner product space is a normed space, it is equipped with a convergence, namely the convergence defined by the norm. This convergence will be called strong convergence.

Definition 3.3.9.

(Strong convergence) A sequence (xn ) of vectors in an inner product space E is called strongly convergent to a vector x in E if xn − x → 0 as n → ∞. The word strong is added in order to distinguish strong convergence from weak convergence.

Definition 3.3.10.

(Weak convergence) A sequence (xn ) of vectors in an inner product space E is called weakly convergent to a vector x in E if xn , y → x, y as n → ∞, for every y ∈ E. The condition in the above definition can be also stated as xn − x, y → 0 as n → ∞, for every y ∈ E. It will be convenient to reserve the notation “xn → x” w for the strong convergence and use “xn → x” to denote weak convergence.

Theorem 3.3.11. A strongly convergent sequence is weakly convergent (to the w same limit), that is, xn → x implies xn → x.

3.3 Hilbert Spaces

103

Proof: Suppose the sequence (xn ) converges strongly to x. This means xn − x → 0 as n → ∞. By Schwarz’s inequality, we have   xn − x, y ≤ xn − x y → 0 as n → ∞, and thus xn − x, y → 0 as n → ∞ for every y ∈ E.




In general, the converse of Theorem 3.3.11 is not true. A suitable example will be given in Section 3.4. For any fixed y in an inner product space E, the mapping ·, y : E → C is a linear functional on E. Theorem 3.3.11 says that such a functional is continuous for every y ∈ E. Obviously, the mapping x, · : E → C is also continuous.

Theorem 3.3.12. If xn → x and yn → y, then xn , yn  → x, y. Proof: If xn → x and yn → y, then

      xn , yn  − x, y ≤ xn , yn  − x, yn  + x, yn  − x, y     = xn − x, yn  + x, yn − y ≤ xn − x yn + x yn − y → 0,

since the sequence (yn ) is bounded.




As an immediate consequence of this theorem, we obtain the following important property of strong convergence: xn → x

implies xn → x .

(3.9)

In the next section it will become clear that Theorem 3.3.12 and property (3.9) need not hold for weak convergence. On the other hand, we have the following theorem. w

Theorem 3.3.13. If xn → x and xn → x , then xn → x. w

Proof: If xn → x, then for all y we have xn , y → x, y

as n → ∞.

Hence, xn , x → x, x = x 2 . Now xn − x 2 = xn − x, xn − x = xn , xn  − xn , x − x, xn  + x, x = xn 2 − 2 Rexn , x + x 2 → x 2 − 2 x 2 + x 2 = 0

104

Chapter 3

Hilbert Spaces and Orthonormal Systems

as n → ∞. Thus, the sequence (xn ) is strongly convergent to x.




The next theorem is often useful in proving weak convergence.

Theorem 3.3.14. Let S be a subset of an inner product space E such that span S is dense in E. If (xn ) is a bounded sequence in E and xn , y → x, y

for every y ∈ S,

w

then xn → x.

Proof: Clearly, if xn , y → x, y for every y ∈ S, then xn , y → x, y for every y ∈ span S. Let z ∈ E and let ε be an arbitrary positive number. Since span S is dense in E, there exists y0 ∈ span S such that z − y0
n0 . 3 Now, for any n > n0 , we have       xn , z − x, z ≤ xn , z − xn , y0  + xn , y0  − x, y0    + x, y0  − x, z ε + x y0 − z 3 ε ε ε <M + +M = ε. 3M 3 3M < xn z − y0 +

w

Since z and ε are arbitrary, we conclude that xn → x.




The following theorem describes an important property of weakly convergent sequences in Hilbert spaces. The proof is not elementary; it is based on the Banach–Steinhaus theorem.

Theorem 3.3.15. Weakly convergent sequences in a Hilbert space are bounded, that is, if (xn ) is a weakly convergent sequence, then there exists a number M such that xn ≤ M for all n ∈ N. Proof: Let (xn ) be a weakly convergent sequence in a Hilbert space H. Define fn (x) = x, xn ,

n ∈ N.

3.4 Orthogonal and Orthonormal Systems

105

Then fn : H → C is a bounded linear functional for every n ∈ N, by Theorem 3.3.11. Since, for every x ∈ H, the sequence (x, xn ) converges, it is bounded, that is, there exists a constant Mx such that     fn (x) = x, xn  ≤ Mx for all n ∈ N. By the Banach–Steinhaus theorem (Theorem 1.5.13), there exists a constant M such that fn ≤ M Since

for all n ∈ N.

    fn (x) = x, xn  ≤ x xn

for all x ∈ H, we have fn ≤ xn . On the other hand,     fn (xn ) = xn , xn  = xn 2 . Consequently, fn = xn , and thus xn ≤ M

for all n ∈ N.




3.4 Orthogonal and Orthonormal Systems By a basis of a vector space E we mean a linearly independentfamily B of vectors from E such that any vector x ∈ E can be written as x = m n=1 λn xn where xn ∈ B and the λn ’s are scalars. In inner product spaces, orthonormal bases are  of much greater importance. Instead of finite combinations m n=1 λn xn , infinite sums are allowed and the condition of linear independence is replaced by orthogonality. One of the immediate advantages of these changes is that, in all important examples, it is possible to describe orthonormal bases. For example, L2 ([−π, π ]) has countable orthonormal bases consisting of simple functions (see Example 3.4.16), while every basis of L2 ([−π, π ]) is uncountable and we can only prove that such a basis exists without being able to describe its elements. In this and the next section, we give all necessary definitions and discuss basic properties of orthonormal bases.

Definition 3.4.1. (Orthogonal and orthonormal systems) A family S of nonzero vectors in an inner product space E is called an orthogonal system if x ⊥ y for any two distinct elements of S. If, in addition, x = 1 for all x ∈ S, then S is called an orthonormal system. Every orthogonal set of nonzero vectors can be normalized: If S is an orx thogonal system, then S1 = { x : x ∈ S} is an orthonormal system.

106

Chapter 3

Hilbert Spaces and Orthonormal Systems

Note that if x is orthogonal to each of y1 , . . . , yn , then x is orthogonal to every linear combination of vectors y1 , . . . , yn . In fact, if y = nk=1 λk yk , then we have ( n ) n   x, y = x, λk yk = λk x, yk  = 0. k=1

k=1

Theorem 3.4.2. Orthogonal systems are linearly independent. n

k=1 αk xk = 0, for some x1 , . . . , xn ∈ S and α1 , . . . , αn ∈ C. Then ( n ) n n n     0= 0, αm xm  = αk x k , αm x m = |αm |2 xm 2 .

Proof: Let S be an orthogonal system. Suppose

m=1

m=1 k=1

m=1

This implies that αm = 0 for each m ∈ N. Thus, x1 , . . . , xn are linearly independent.


Definition 3.4.3.

(Orthonormal sequence) A sequence of vectors which constitutes an orthonormal system is called an orthonormal sequence.

In applications, it is often convenient to use sequences indexed by the set of all integers, Z. The condition of orthonormality of a sequence (xn ) can be expressed in terms of the Kronecker delta symbol (Leopold Kronecker (1823– 1891)): * (3.10) xm , xn  = δmn = 0 if m = n, 1 if m = n.

Example 3.4.4. For en = (0, . . . , 0, 1, 0, . . .) with 1 in the nth position, the set S = {e1 , e2 , . . .} is an orthonormal system in l2 .




inx

Example 3.4.5. Let ϕn (x) = √e 2π , n ∈ Z. The set {ϕn : n ∈ Z} is an orthonormal

system in L2 ([−π, π ]). Indeed, for m = n, we have  π 1 eπ i(m−n) − e−π i(m−n) = 0. ϕm , ϕn  = ei(m−n)x dx = 2π −π 2π i(m − n) On the other hand, ϕn , ϕn  =

1 2π



π −π

ei(n−n)x dx = 1.

Thus, ϕm , ϕn  = δmn for every pair of integers m and n.




107

3.4 Orthogonal and Orthonormal Systems

Example 3.4.6. The Legendre polynomials (Adrien-Marie Legendre (1752– 1833)) defined by P0 (x) = 1, Pn (x) =

1 dn 2 (x − 1)n , 2n n! dxn

n = 1, 2, 3, . . .

(3.11)

form an orthogonal system in L2 ([−1, 1]). For convenience, we write (x2 − 1)n = pn (x) so that  1  1 1 m Pn (x)xm dx = n p(n) (3.12) n (x)x dx. 2 n! −1 −1 We evaluate this integral for m < n by recursion. First we note that p(k) n (x) = 0 for x = ±1 and k = 0, 1, 2, . . . , n − 1. Hence, by integration by parts in (3.12), we get  1  1 (n) m pn (x)x dx = −m p(n−1) (x)xm−1 dx. n −1

−1

Repeated application of this operation ultimately leads to  1  1 p(n−m) (x)dx = (−1)m m! p(n−m−1) (x) −1 = 0 (m < n). (−1)m m! n n −1

Consequently,



1 −1

Pn (x)xm dx = 0,

for m < n.

Since Pm is a polynomial of degree m, it follows that  1 Pn , Pm  = Pn (x)Pm (x)dx = 0, for n = m. −1

(3.13)

(3.14)

This proves orthogonality of the Legendre polynomials. To obtain an orthonormal system from the Legendre polynomials, we need to evaluate the norm of Pn in L2 ([−1, 1]): + 1 2 Pn = Pn (x) dx. −1

By repeated integration by parts, we first obtain  1  1   2 n (1 − x)n (1 + x)n dx 1 − x dx = −1

−1

108

Chapter 3

Hilbert Spaces and Orthonormal Systems



1

=

n n+1

=

n(n − 1) · · · 2 · 1 (n + 1)(n + 2) · · · 2n

=

−1

(1 − x)n−1 (1 + x)n+1 dx = · · · 

1 −1

(1 + x)2n dx

2 2n+1

(n!) 2 . (2n)!(2n + 1)

(3.15)

A similar procedure gives 

2 p(n) n (x) dx = 0 −

1 −1



1 −1

= ···

p(n−1) (x)p(n+1) (x)dx n n



= (−1)n  = (2n)!

1 −1 1

−1

pn (x)p(2n) n (x)dx

(1 − x)n (1 + x)n dx,

(3.16)

where we have used the fact that the 2nth derivative of pn (x) = (x2 − 1)n is the same as the derivative of the term of exponent 2n. The 2nth derivatives of all the other terms of the sum are zero. Now from (3.11), (3.15), and (3.16) we obtain  1 2  1 (n!)2 22n+1 2 Pn (x) dx = n 2 (2n)! = . (3.17) (2 n!) (2n)!(2n + 1) 2n + 1 −1

Thus, the polynomials

n + 12 Pn (x) form an orthonormal system in

L2 ([−1, 1]).




Example 3.4.7. Denote by Hn the Hermite polynomial (Charles Hermite (1822–1901)) of degree n, that is, Hn (x) = (−1)n ex The functions ϕn (x) = e−x inner product

2 /2

2

dn −x2 e . dxn

(3.18)

Hn (x) form an orthogonal system in L2 (R). The 

ϕn , ϕm  = (−1)n+m

∞ −∞

ex

2

dn −x2 dm −x2 e e dx dxn dxm

can be evaluated by integration by parts, which gives

109

3.4 Orthogonal and Orthonormal Systems

dn −x2 dm−1 −x2 ∞ ϕn , ϕm  = e e e dxn dxm−1 −∞ - m−1  ∞ , n d x2 d −x2 d 2 e − e e−x dx, n m−1 dx dx dx −∞ ,

(−1)

n+m

x2

(3.19)

and hence all terms under the differential sign contain the factor e−x . Since for any k ∈ N, we have 2

xk e−x → 0 as x → ∞, 2

the first term in (3.19) vanishes. Therefore, repeated integration by parts leads to ϕn , ϕm  = 0 for n = m.

(3.20)

To obtain an orthonormal system, we evaluate the norm:  ϕn =

∞

2

−∞

e

−x2

2  Hn (x) dx =



∞ −∞

e

−x2

,

dn −x2 e e dxn x2

-2 dx.

Integration by parts n times yields  ϕn 2 = (−1)n

∞ −∞

e−x

2

, dn x2 dn −x2 e e dx. dxn dxn

Since Hn is a polynomial of degree n, direct differentiation gives ex and

2

dn −x2 e = (−2x)n + · · · dxn

,  dn  dn x2 dn −x2 n e + · · · = (−1)n 2n n!. e = (−2x) dxn dxn dxn

Consequently,

 ϕn 2 = 2n n!

∞

√ 2 e−x dx = 2n n! π.

−∞

(3.21)

Thus, the functions ψn (x) =

1

−x √ e π

2n n!

form an orthonormal system in L2 (R).

2 /2

Hn (x)


110

Chapter 3

Hilbert Spaces and Orthonormal Systems

In the preceding examples, the original sequence of functions is orthogonal but not orthonormal. Although the calculations involved might be complicated, it is always possible to normalize the functions and obtain an orthonormal sequence. It turns out that the same is possible if the original sequence of functions (or, in general, a sequence of vectors in an inner product space) is linearly independent, not necessarily orthogonal. The method of transforming such a sequence into an orthonormal sequence is called the Gram– Schmidt orthonormalization process (Jorgen Pedersen Gram (1850–1916), Erhard Schmidt (1876–1959)). The process can be described as follows: Given a sequence (yn ) of linearly independent vectors in an inner product space, define sequences (wn ) and (xn ) inductively by w1 = y1 , wk = yk −

w1 , w1 wk xk = , wk

x1 = k−1 

yk , xn xn ,

n=1

for k = 2, 3, . . . .

The sequence (wn ) is orthogonal. We show that by induction. First note that w2 , w1  = y2 − y2 , x1 x1 , y1  = y2 , y1  − y2 , x1 x1 , y1  = y2 , y1  −

y2 , y1 y1 , y1  = 0. y1 2

Assume now that w1 , . . . , wk−1 are orthogonal. Then, for any m < k, k−1 yk , wn wn , wm  wk , wm  = yk , wm  − n=1 wm 2 = yk , wm  −

yk , wm wm , wm  = 0. wm 2

Therefore, vectors w1 , . . . , wk are orthogonal. It follows, by induction, that the sequence (wn ) is orthogonal, and thus (xn ) is orthonormal. It is easy to check that any linear combination of vectors x1 , . . . , xn is also a linear combination of y1 , . . . , yn and vice versa. In other words, span{x1 , . . . , xn } = span{y1 , . . . , yn } for every n ∈ N. In Section 3.3 we proved that the Pythagorean formula holds for any pair of orthogonal vectors in an inner product space. It turns out that it can be generalized to any finite number of orthogonal vectors.

Theorem 3.4.8. (Pythagorean formula) If x1 , . . . , xn are orthogonal vectors in an inner product space, then  n 2 n      xk  = xk 2 . (3.22)    k=1

k=1

111

3.4 Orthogonal and Orthonormal Systems

Proof: If x1 ⊥ x2 , then x1 + x2 2 = x1 2 + x2 2 , by (3.8). Thus the theorem is true for n = 2. Assume now that the (3.22) holds for n − 1, that is,  n−1 2 n−1      xk  = xk 2 .    k=1

k=1

n−1

Set x = k=1 xk and y = xn . Since x ⊥ y, we have  n 2 n−1 n       xk  = x + y 2 = x 2 + y 2 = xk 2 + xn 2 = xk 2 .    k=1

k=1

k=1

This proves the theorem.




The Pythagorean formula will help us prove the following important property of orthonormal sets attributed to Friedrich Wilhelm Bessel (1784–1846).

Theorem 3.4.9. (Bessel’s equality and inequality) Let x1 , . . . , xn be an orthonormal set of vectors in an inner product space E. Then, for every x ∈ E, we have  2 n n         x, xk 2 x, xk xk  = x 2 − (3.23) x −   k=1

k=1

and n    x, xk 2 ≤ x 2 .

(3.24)

k=1

Proof: In view of the Pythagorean formula (3.22), we have  n 2 n n       αk x k  = αk xk 2 = |αk |2    k=1

k=1

k=1

for arbitrary complex numbers α1 , . . . , αn . Hence  2 ( ) n n n        αk x k  = x − αk x k , x − αk x k x −   k=1 k=1 k=1 ) ( n ( n ) n    = x 2 − x, αk x k − αk x k , x + |αk |2 xk 2 k=1

= x 2 − = x 2 −

n 

αk x, xk  −

k=1 n 

k=1 n 

k=1 n 

k=1

k=1

  x, xk 2 +

αk x, xk  +

k=1 n 

αk αk

k=1

  x, xk  − αk 2 .

(3.25)

112

Chapter 3

Hilbert Spaces and Orthonormal Systems

In particular, if αk = x, xk , this results yields (3.23). From (3.23) it follows that n    2 x, xk 2 , 0 ≤ x − k=1

which gives (3.24).




Note that expression (3.25)  is minimized by taking αk = x, xk . This choice of αk ’s minimizes x − nk=1 αk xk , and thus it provides the best approximation of x by a linear combination of vectors x1 , . . . , xn . This property of orthonormal systems is of fundamental importance for many approximation techniques. If (xn ) is an orthonormal sequence, then by letting n → ∞ in (3.24), we obtain ∞    x, xk 2 ≤ x 2 (3.26) k=1

and consequently, lim x, xn  = 0.

n→∞

Therefore, orthonormal sequences are weakly convergent to zero. On the other hand, since xn = 1 for all n ∈ N, orthonormal sequences are not strongly convergent.  2 Inequality (3.26) implies that the series ∞ k=1 |x, xk | converges for every x ∈ E. In other words, the sequence (x, xn ) is an element of l2 . We can say that an orthonormal sequence in E induces a mapping from E into l2 . The expansion x∼

∞  x, xn xn

(3.27)

n=1

is called a generalized Fourier series of x (Jean Baptiste Joseph Fourier (1768– 1830)). The scalars αn = x, xn  are called the generalized Fourier coefficients of x with respect to the orthonormal sequence (xn ). As mentioned earlier, this set of coefficients gives the best approximation for any finite set of xn ’s. In general, we do not know whether the series in (3.27) is convergent. However, as the next theorem shows, completeness of the space ensures the convergence.

Theorem 3.4.10. Let (xn ) be an orthonormal sequence in a Hilbert space H, and ∞ let (αn ) be a sequence of complex numbers. Then the series n=1 αn xn converges ∞ if and only if n=1 |αn |2 < ∞ and in that case ∞ 2 ∞      αn x n  = |αn |2 . (3.28)    n=1

n=1

3.4 Orthogonal and Orthonormal Systems

113

Proof: For every m > k > 0, we have  m 2 m      αn x n  = |αn |2 ,    n=k

(3.29)

n=k

 2 by the Pythagorean formula (3.22).If ∞ n=1 |αn | < ∞, then from (3.29) m we obtain that the sequence sm =  n=1 αn xn is a Cauchy sequence. This implies convergence of the series ∞ α x , because of completeness of H. ∞ n=1 n n Conversely, if theseries n=1 αn xn converges, then (3.29) implies ∞ 2 the convergence of |α | , because the sequence of numbers σm = n n=1 m 2 |α | is a Cauchy sequence in R. n n=1 To obtain (3.28), it is enough to take k = 1 and let m → ∞ in (3.29).
theorem and (3.26) imply that in a Hilbert space H the series This ∞ x, xn xn converges for every x ∈ H. However, it can happen that it conn=1 verges to an element different from x.

Example 3.4.11. Let H = L2 ([−π, π ]), and let xn (t) =

sin nt for n = 1, 2, . . . . The sequence (xn ) is an orthonormal set in H. On the other hand, for x(t) = cos t, we have √1 π

 π ∞ ∞ ,   sin nt 1 x, xn xn (t) = cos t sin ntdt √ √ π −π π n=1 n=1 =

∞ 

0 · sin nt = 0 = cos t.




n=1

Definition 3.4.12.

(Complete orthonormal sequence) An orthonormal sequence (xn ) in an inner product space E is said to be complete if for every x ∈ E we have x=

∞ 

x, xn xn .

(3.30)

n=1

It is important to remember that since the right-hand side of (3.30) is an infinite series, the equality means   n      lim x − x, xk xk  = 0, n→∞  k=1

114

Chapter 3

Hilbert Spaces and Orthonormal Systems

where · is the norm in E. For example, if E = L2 ([−π, π ]) and (fn ) is an orthonormal sequence in E, then by f=

∞  f , fn fn , n=1

we mean

2  n      αk fk (t) dt = 0, lim f (t) − n→∞ −π   

π

 where αk =

k=1

π −π

f (t)fk (t)dt.

This, in general,  does not imply pointwise convergence, so we cannot claim that f (x) = ∞ n=1 αn fn (x).

Definition 3.4.13.

(Orthonormal basis) An orthonormal system B in an inner product space E is called an orthonormal basis if every x ∈ E has a unique representation x=

∞ 

αn x n ,

n=1

where αn ∈ C and xn ’s are distinct elements of B. Note that a complete orthonormal sequence (xn ) is an orthonormal basis. It suffices to show the uniqueness. Indeed, if x=

∞ 

αn x n

and

x=

n=1

∞ 

βn xn ,

n=1

then

∞ 2  ∞ 2 ∞ ∞           αn x n − βn xn  =  (αn − βn )xn  = |αn − βn |2 0 = x − x =      2

n=1

n=1

n=1

n=1

by Theorem 3.4.10. This means that αn = βn for all n ∈ N, proving the uniqueness. If (xn ) is a complete orthonormal sequence in an inner product space E, then the set  n   span{x1 , x2 , . . .} = αk xk : n ∈ N, α1 , . . . , αk ∈ C k=1

is dense in E.

3.4 Orthogonal and Orthonormal Systems

115

The following two theorems give important characterizations of complete orthonormal sequences in Hilbert spaces.

Theorem 3.4.14. An orthonormal sequence (xn ) in a Hilbert space H is complete if and only if x, xn  = 0 for all n ∈ N implies x = 0. Proof: Suppose (xn ) is a complete orthonormal sequence in H. Then for every x ∈ H, we have x=

∞  x, xn xn . n=1

Thus, if x, xn  = 0 for every n ∈ N, then x = 0. Conversely, suppose x, xn  = 0 for all n ∈ N implies x = 0. Let x be an element of H. Define ∞  x, xn xn . y= n=1

The sum y exists in H by (3.26) and Theorem 3.4.10. Since, for every n ∈ N, ) (∞  x, xk xk , xn x − y, xn  = x, xn  − k=1

= x, xn  −

∞ 

x, xk xk , xn 

k=1

= x, xn  − x, xn  = 0,  we have x − y = 0, and hence x = ∞ n=1 x, xn xn .




The formula in the following theorem, known as Parseval’s formula (MarcAntoine Parseval des Chênes (1755–1836)), can be interpreted as an extension of the Pythagorean formula to infinite sums.

Theorem 3.4.15. (Parseval’s formula) An orthonormal sequence (xn ) in a Hilbert space H is complete if and only if ∞    x, xn 2 x = 2

(3.31)

n=1

for every x ∈ H.

Proof: Let x ∈ H. By (3.23), for every n ∈ N, we have 2  n n         x, xk 2 . x, xk xk  = x 2 − x −   k=1

k=1

(3.32)

116

Chapter 3

Hilbert Spaces and Orthonormal Systems

If (xn ) is a complete sequence, then the expression on the left in (3.32) converges to zero as n → ∞. Hence, & ' n    2 x, xk  = 0. lim x 2 − n→∞

k=1

Therefore, (3.31) holds. Conversely, if (3.31) holds, then the expression on the right in (3.32) converges to zero as n → ∞, and thus  2 n      x, xk xk  = 0. lim x − n→∞  k=1

This proves that the sequence (xn ) is complete.




Example 3.4.16. The orthonormal system einx ϕn (x) = √ , 2π

n = 0, ±1, ±2, . . . ,

given in Example 3.4.5, is complete in L2 ([−π, π ]). The proof of completeness is not simple. It will be discussed in Section 3.5. A simple change of scale allows us to represent a function f ∈ L2 ([0, a]) in the form ∞  f (x) = βn e2nπ ix/a n=−∞

where βn =

1 a



a

f (t)e−2nπ it/a dt.




0

Example 3.4.17. The sequence of functions 1 cos x sin x cos 2x sin 2x √ , √ , √ , √ , √ ,... π π π π 2π is a complete orthonormal system in L2 ([−π, π ]). The orthogonality follows from the following identities by simple integration: 2 cos nx cos mx = cos(n + m)x + cos(n − m)x, 2 sin nx sin mx = cos(n − m)x − cos(n + m)x, 2 cos nx sin mx = sin(n + m)x − sin(n − m)x.

3.4 Orthogonal and Orthonormal Systems

Since



π

 cos nx dx =

π

2

−π

−π

117

sin2 mx dx = π,

the sequence is also orthonormal. Completeness follows from completeness of the sequence in Example 3.4.16, in view of the following identities: e0 = 1 and einx = (cos nx + i sin nx).




Example 3.4.18. Each of the following two sequences of functions is a complete orthonormal system in L2 ([0, π ]): . . . 1 2 2 2 cos x, cos 2x, cos 3x, . . . , √ , π π π π . . . 2 2 2 sin x, sin 2x, sin 3x, . . . . π π π




Example 3.4.19. (Rademacher functions and Walsh functions) Rademacher functions R(m, x) (Hans Rademacher (1892–1969)) can be introduced in many different ways. We will use the definition based on the sine function:   R(m, x) = sgn sin(2m π x) ,

m = 0, 1, 2, . . . , x ∈ [0, 1],

where sgn denotes the signum function defined by  sgn(x) =

1 0 −1

if x > 0, if x = 0, if x < 0.

Rademacher functions form an orthonormal system in L2 ([0, 1]). Obviously,  1   R(m, x)2 dx = 1 for all m. 0

To show that for m = n, we have 

1

R(m, x)R(n, x)dx = 0,

0

b first notice that a R(m, x)dx = 0 whenever 2m (b − a) is an even number. Thus, for m > n ≥ 0, we have

118 

Chapter 3 1

Hilbert Spaces and Orthonormal Systems



1

R(m, x)R(n, x)dx =

R(m, x)R(n, x)dx

0

0 2   n

=

k=1 2  n

=

k=1

k/2n

R(m, x)R(n, x)dx (k−1)/2n

    k/2n 2k − 1 sgn R n, R(m, x)dx = 0, 2 (k−1)/2n

because all the integrals equal 0. The sequence of Rademacher functions is not complete. Indeed, consider the function  1   0 if 0 ≤ x < 4 , f (x) = 1 if 14 ≤ x ≤ 34 ,   0 if 34 < x ≤ 1. Then 

1 0

1 R(0, x)f (x)dx = 2



1

and

R(m, x)f (x)dx = 0 for m ≥ 1,

0

but f (x) = 12 R(0, x). Rademacher functions can be used to construct Walsh functions (John Walsh (1786–1847)), which form a complete orthonormal system. Walsh functions are denoted by W(m, x), m = 0, 1, 2, . . . . For m = 0, we set W(0, x) = 1. For other values of m, we first represent m as a binary number, that is, m=

n 

2k−1 ak = a1 + 21 a2 + 22 a3 + · · · + 2n−1 an ,

k=1

where a1 , a2 , . . . , an = 0 or 1. Then we define W(m, x) =

n  a   a  a a R(k, x) k = R(1, x) 1 R(2, x) 2 · · · R(n, x) n k=1

(where (R(m, x))0 ≡ 1). For instance, since 53 is written as 110101 in binary form, we have W(53, x) = R(1, x)R(3, x)R(5, x)R(6, x). Clearly, we have R(n, x) = W(2n−1 , x), Figure 3.2 shows several Walsh functions.

n ∈ N.


3.4 Orthogonal and Orthonormal Systems

119

Figure 3.2 Walsh functions.

Definition 3.4.20.

(Separable spaces) A Hilbert space is called separable if it contains a complete orthonormal sequence. Finite dimensional Hilbert spaces are considered separable.

Example 3.4.21. Space L2 ([−π, π ]) is separable. Example 3.4.16 shows a complete orthonormal sequence in L2 ([−π, π ]).




Example 3.4.22. Space l2 is separable.




Example 3.4.23. (Nonseparable Hilbert space) Let H be the space of all complex-valued functions defined on R, which vanish everywhere except a countable number of points in R and such that   f (x)2 < ∞. f (x)=0

120

Chapter 3

Hilbert Spaces and Orthonormal Systems

The inner product in H can be defined as  f , g = f (x)g(x). f (x)g(x)=0

This space is not separable because for any sequence of functions fn ∈ H, there are nonzero functions f such that f , fn  = 0 for all n ∈ N.
Throughout this book, we will consider only separable Hilbert spaces. Recall that a set S in a Banach space E is called dense in E if every element of E can be approximated by a sequence of elements of S. More precisely, for every x ∈ E there exist xn ∈ S such that x − xn → 0 as n → ∞.

Theorem 3.4.24. Every separable Hilbert space contains a countable dense subset.

Proof: Let (xn ) be a complete orthonormal sequence in a Hilbert space H. The set
 S = (α1 + iβ1 )x1 + · · · + (αn + iβn )xn : α1 , . . . , αn , β1 , . . . , βn ∈ Q, n ∈ N is obviously countable. Since, for every x ∈ H,   n      x, xk xk − x → 0 as n → ∞,   k=1

S is dense in H.




The statement in Theorem 3.4.24 is often used as a definition of separability.

Theorem 3.4.25. Every orthogonal set in a separable Hilbert space is countable. Proof: Let S be an orthogonal set in a separable Hilbert space H, and let S1 be the set of normalized vectors from S, that is, S1 = {x/ x : x ∈ S}. For any distinct x, y ∈ S1 we have x − y 2 = x − y, x − y = x, x − x, y − y, x + y, y = 1 − 0 − 0 + 1 (by the orthogonality) = 2. √ This means that the distance between any two distinct elements of S is 2. 1 √ Now consider the collection of 1/ 2-neighborhoods about every element of S1 . Clearly, no two of these neighborhoods can have a common point. Since every dense subset of H must have at least one point in every

3.4 Orthogonal and Orthonormal Systems

121

neighborhood and H has a countable dense subset, S1 has to be countable. Thus, S is countable, proving the theorem.


Definition 3.4.26.

(Hilbert space isomorphism) A Hilbert space H1 is said to be isomorphic to a Hilbert space H2 if there exists a one-to-one linear mapping T from H1 onto H2 such that $ % (3.33) T(x), T(y) = x, y for every x, y ∈ H1 . Such a mapping T is called a Hilbert space isomorphism of H1 onto H2 . Note that (3.33) implies T = 1, because T(x) = x for every x ∈ H1 .

Theorem 3.4.27. Let H be a separable Hilbert space. (a) If H is infinite dimensional, then it is isomorphic to l2 ; (b) If dim H = N, then it is isomorphic to CN .

Proof: Let (xn ) be a complete orthonormal sequence in H. If H is infinite dimensional, then (xn ) is an infinite sequence. Let x be an element of H. Define T(x) = (αn ), where αn = x, xn , n = 1, 2, . . . . By Theorem 3.4.10, T is a one-to-one mapping from H onto l2 . It is clearly a linear mapping. Moreover, for αn = x, xn  and βn = y, xn , x, y ∈ H, n ∈ N, we have $ % $ % T(x), T(y) = (αn ), (βn ) =

∞  n=1

αn β n =

∞ 

x, xn y, xn 

n=1

) ( ∞ ∞   $ % y, xn xn = x, y. x, y, xn xn = x, = n=1

Thus, T is an isomorphism from H onto l2 . The proof of (b) is left as an exercise.

n=1




It is easy to check that isomorphism of Hilbert spaces is an equivalence relation. Since any infinite dimensional separable Hilbert space is isomorphic to l2 , it follows that any two such spaces are isomorphic. The same is true for real Hilbert spaces; any real infinite dimensional separable Hilbert space is isomorphic to the real space l2 . In some sense, there is only one real and one complex infinite dimensional separable Hilbert space.

122

Chapter 3

Hilbert Spaces and Orthonormal Systems

3.5 Trigonometric Fourier Series In this section we prove that the sequence einx ϕn (x) = √ , 2π

n = 0, ±1, ±2, . . . ,

is a complete orthonormal sequence in L2 ([−π, π ]). The orthogonality has been established in Example 3.4.5. The proof of completeness is much more complicated. For the purpose of this proof it will be convenient to identify elements of L1 ([−π, π ]) with 2π -periodic locally integrable functions on R, because then we have  π −x  π  π f (t)dt = f (t)dt = f (t − x)dt −π

−π −x

−π

for any f ∈ L1 ([−π, π ]) and any x ∈ R. Let f ∈ L1 ([−π, π ]) and fn =

n 

f , ϕk ϕk ,

n = 0, 1, 2, . . . .

k=−n

Then fn (x) =

 π  π n n   1 1 f (t)e−ikt dt eikx = f (t)eik(x−t) dt. 2π −π 2π −π

k=−n

k=−n

We intend to show that for every f ∈ L1 ([−π, π ]), we have f0 + f1 + · · · + fn =f n→∞ n+1 lim

in the L1 ([−π, π ]) norm. First observe that  n   |k| f0 (x) + f1 (x) + · · · + fn (x) 1− = f , ϕk ϕk (x) n+1 n+1 k=−n

  π n  |k| 1 f (t)e−ikt dt eikx 1− = 2π n + 1 −π k=−n

1 = 2π



   n  |k| ik(x−t) f (t) 1− dt. e n+1 −π π

k=−n

(3.34)

3.5 Trigonometric Fourier Series

123

Lemma 3.5.1. For every n ∈ N and x ∈ R we have n   k=−n

 (n+1)x 1 sin2 2 |k| . 1− eikx = n+1 n + 1 sin2 2x

Proof: First note that sin2

1 1 1 x 1 = (1 − cos x) = − e−ix + − eix . 2 2 4 2 4

Then routine calculations give   n   1 −ix 1 1 ix  |k| − e + − e 1− eikx 4 2 4 n+1 k=−n   1 1 −i(n+1)x 1 1 i(n+1)x = + − e , − e n+1 4 2 4 proving the lemma.




Definition 3.5.2. (Summability kernel) By a summability kernel, we mean a sequence (κn ) of 2π -periodic continuous functions satisfying:  π κn (t)dt = 2π for all n ∈ N, (3.35) −π  π   κn (t)dt ≤ M for some M and all n ∈ N, (3.36) −π

lim



n→∞ δ

2π −δ 

 κn (t)dt = 0

for all δ ∈ (0, π ).

(3.37)

Lemma 3.5.3. The sequence of functions  n   |k| eikt Kn (t) = 1− n+1 k=−n

is a summability kernel.  π ikt π Proof: Since −π e dt = 2π if k = 0 and −π eikt dt = 0 for any other integer k, we obtain  π  π n   |k| Kn (t)dt = eikt dt = 2π. 1− n + 1 −π −π k=−n

124

Chapter 3

Hilbert Spaces and Orthonormal Systems

From Lemma 3.5.1, it follows that Kn ≥ 0 and thus  π  π   Kn (t)dt = Kn (t)dt = 2π. −π

−π

Finally, let δ ∈ (0, π ). For t ∈ (δ, 2π − δ), we have sin 2t ≥ sin 2δ and hence Kn (t) = Thus,



(n+1)x 2 x 2

1 sin2 n + 1 sin2

2π −δ

Kn (t)dt ≤

δ

≤

1 . (n + 1) sin2 2δ

2π . (n + 1) sin2 2δ

For a fixed δ the right-hand side tends to 0 as n → ∞.




The kernel in Lemma 3.5.3 is called the Fejér’s kernel (Lipót Fejér (1880– 1959)). The reason for introducing here the notion of a summability kernel is the following important theorem.

Theorem 3.5.4. Let (κn ) be a summability kernel and let f ∈ L1 ([−π, π ]). Then 1 n→∞ 2π lim



π −π

κn (t)f (x − t)dt = f (x)

in the L1 ([−π, π ]) norm, that is,   π  π  1   lim κn (t)f (x − t)dt − f (x)dx = 0.  n→∞ −π 2π −π

(3.38)

Proof: By (3.35), we have 1 f (x) = 2π



π

−π

κn (t)f (x)dt

and hence  π  π   1 1 κn (t)f (x − t)dt − f (x) = κn (t) f (x − t) − f (x) dt. 2π −π 2π −π Consequently,   π  π   1  κn (t)f (x − t)dt − f (x)dx  2π −π −π   π  π  1     = κn (t) f (x − t) − f (x) dt dx  −π 2π −π

125

3.5 Trigonometric Fourier Series    π  δ  1     = κn (t) f (x − t) − f (x) dt dx  −π 2π −δ   π  2π −δ  1     κn (t) f (x − t) − f (x) dt dx +  −π 2π δ  π δ    1 κn (t) f (x − t) − f (x) dt dx ≤ 2π −π −δ  π  2π −δ    1 κn (t) f (x − t) − f (x) dt dx. + 2π −π δ

Since 1 2π





π −π

δ

−δ

  κn (t) f (x − t) − f (x) dt dx

 π   π     1 f (x − t) − f (x)dx κn (t)dt, ≤ max 2π |t|≤δ −π −π

and since, by Theorem 2.4.3, we have 

π

lim

t→0 −π

  f (x − t) − f (x)dx = 0,

for an arbitrary ε > 0 we can find δ > 0 small enough to ensure that 1 2π





π

−π

δ −δ

  κn (t) f (x − t) − f (x) dt dx < ε

(3.39)

π for all n ∈ N, since, by (3.36), −π |κn (t)|dt is bounded. On the other hand, we have 

π −π



2π −δ 

  κn (t) f (x − t) − f (x) dt dx

δ

  ≤ max |t|≤π

 ≤2

π

−π

π −π

  f (x − t) − f (x)dx

  f (x)dx





2π −δ 

 κn (t)dt

δ 2π −δ 

 κn (t)dt → 0 as n → ∞,

(3.40)

δ

by (3.37). Combining (3.39) and (3.40), we obtain the proof of (3.38).




Theorem 3.5.5. If f ∈ L1 ([−π, π ]) and f , ϕn  = 0 for all integers n, then f = 0 a.e.

126

Chapter 3

Hilbert Spaces and Orthonormal Systems

Proof: If



π −π

f (t)e−int dt = 0,

for all integers n, then  π n  1 f (t)eik(x−t) dt = 0, fn (x) = 2π −π k=−n

and consequently, 1 f0 (x) + f1 (x) + · · · + fn (x) = n+1 2π



  n   |k| ik(x−t) e dt f (t) 1− n+1 −π 

π

k=−n

= 0. On the other hand, since f and all the functions eikx are 2π -periodic, we have   n    π  1 |k| eik(x−t) dt f (t) 1− 2π −π n+1 k=−n  n     π  |k| 1 f (x − t) 1− = eikt dt 2π −π n+1 k=−n

and hence, by Theorem 3.5.4 and Lemma 3.5.3, lim

n→∞

f0 + f1 + · · · + fn =f n+1

in the L1 ([−π, π ]) norm. Therefore, f = 0 a.e., proving the theorem.




Theorem 3.5.6. The sequence einx ϕn (x) = √ , 2π

n = 0, ±1, ±2, . . . ,

is complete.

Proof: If f ∈ L2 ([−π, π ]), then f ∈ L1 ([−π, π ]). Thus, by Theorem 3.5.5,

if f , ϕn  = 0 for all integers n, then f = 0 a.e., that is, f = 0 in L2 ([−π, π ]). This proves completeness of the sequence by Theorem 3.4.14.
Theorem 3.5.6 implies that, for every f ∈ L2 ([−π, π ]), we have f=

∞  k=−∞

αn ϕ n

(3.41)

3.6 Orthogonal Complements and Projections

where einx 1 and αn = √ ϕn (x) = √ 2π 2π In this case Parseval’s formula yields  f 2 =

π −π



π −π

127

f (t)e−ikt dt.

∞    f (x)2 dx = |αn |2 . n=−∞

The series (3.41) is called the Fourier series of f and the numbers αn are called the Fourier coefficients of f . It is important to remember that, in general, (3.41) does not imply pointwise convergence. The problem of pointwise convergence of Fourier series is much more difficult. In 1966, L. Carleson proved that Fourier series of functions in L2 ([−π, π ]) converge almost everywhere.

3.6 Orthogonal Complements and Projections By a subspace of an inner product space E, we mean a vector subspace of E. A subspace of an inner product space is an inner product space itself. A closed subspace of a Hilbert space is a Hilbert space, because a closed subspace of a complete normed space is complete.

Definition 3.6.1. (Orthogonal complement) Let S be a nonempty subset of an inner product space E. An element x ∈ E is said to be orthogonal to S, denoted by x ⊥ S, if x, y = 0 for every y ∈ S. The set of all elements of E orthogonal to S, denoted by S⊥ , is called the orthogonal complement of S. In symbols: S⊥ = {x ∈ E, x ⊥ S}.

The orthogonal complement of S⊥ is denoted by S⊥⊥ = (S⊥ )⊥ . The notation S⊥ is imprecise, because the complement depends on E. If E1 and E2 are two different inner product spaces and S ⊂ E1 ∩ E2 , then obviously the complement of S in E1 is not the same as the complement in E2 . The notation S⊥ is used when it is clear what the inner product space is. For instance, in Definition 3.6.1 the underlying space is E and thus all complements are with respect to E. If x ⊥ y for every y ∈ E, then x = 0. Thus E⊥ = {0}. Similarly, {0}⊥ = H. Two subsets A and B of an inner product space are said to be orthogonal if x ⊥ y for every x ∈ A and y ∈ B. This is denoted by A ⊥ B. Note that if A ⊥ B, then A ∩ B = {0} or ∅.

128

Chapter 3

Hilbert Spaces and Orthonormal Systems

Theorem 3.6.2. For any subset S of an inner product space E, the set S⊥ is a closed subspace of E.

Proof: If α, β ∈ C and x, y ∈ S⊥ , then αx + βy, z = αx, z + βy, z = 0 for every z ∈ S. Thus, S⊥ is a vector subspace of E. We next prove that S⊥ is closed. Let (xn ) ∈ S⊥ and xn → x for some x ∈ E. From continuity of the inner product, we have 0 / x, y = lim xn , y = lim xn , y = 0, n→∞

n→∞

for every y ∈ S. This shows that x ∈ S⊥ , and thus, S⊥ is closed.




This theorem implies that S⊥ is a Hilbert space for any subset S of a Hilbert space H. Note that S does not have to be a vector space. Since S ⊥ S⊥ , we have S ∩ S⊥ = {0} or S ∩ S⊥ = ∅.

Definition 3.6.3.

(Convex set) A set U in a vector space is called convex if for any x, y ∈ U and α ∈ (0, 1) we have αx + (1 − α)y ∈ U . Note that a vector subspace is a convex set. The following theorem, concerning the minimization of the norm, is of fundamental importance in approximation theory.

Theorem 3.6.4. (The closest point property) Let S be a closed convex subset of a Hilbert space H. For every point x ∈ H there exists a unique point y ∈ S such that x − y = inf x − z . z∈S

Proof: Let (yn ) be a sequence in S such that lim x − yn = inf x − z .

n→∞

z∈S

Denote d = infz∈S x − z . Since 12 (ym + yn ) ∈ S, we have     x − 1 (ym + yn ) ≥ d,   2

for all m, n ∈ N.

(3.42)

3.6 Orthogonal Complements and Projections

129

Moreover, by the parallelogram law (3.5),  2  2     1 1 2 2    ym − yn = 4 x − 2 (ym + yn ) + ym − yn − 4x − 2 (ym + yn )  2  2 = (x − ym ) + (x − yn ) + (x − ym ) − (x − yn )  2   1  − 4x − (ym + yn )  2 2      1  (y = 2 x − ym 2 + x − yn 2 − 4 + y ) x − m n  .  2 Since and

  2 x − ym 2 + x − yn 2 → 4d2 ,

as m, n → ∞,

 2   x − 1 (ym + yn ) ≥ d2 ,   2

we have ym − yn 2 → 0, as m, n → ∞. Thus, (yn ) is a Cauchy sequence. Since H is complete and S is closed, the limit limn→∞ yn = y exists and y ∈ S. From the continuity of the norm, we obtain     x − y = x − lim yn  = lim x − yn = d. n→∞

n→∞

We have proved that there exists a point in S satisfying (3.42). It remains to prove the uniqueness. Suppose there is another point y1 in S satisfying (3.42). Then, since 12 (y + y1 ) ∈ S, we have  2  y + y1    ≤ 0. y − y1 = 4d − 4x − 2  2

2

This can only happen if y = y1 .




Theorem 3.6.4 gives an existence and uniqueness result, which is crucial for optimization problems. However, it does not tell us how to find that optimal point. The characterization of the optimal point in the case of a real inner product space, stated in the following theorem, is often useful in such problems.

Theorem 3.6.5. Let S be a convex subset of a real inner product space E, y ∈ S and let x ∈ E. Then the following conditions are equivalent: (a) x − y = infz∈S x − z , (b) x − y, z − y ≤ 0 for all z ∈ S.

130

Chapter 3

Hilbert Spaces and Orthonormal Systems

Figure 3.3 Geometric interpretation of Theorem 3.6.5 in R2 .

Proof: Let z ∈ S. Since S is convex, λz + (1 − λ)y ∈ S for every λ ∈ (0, 1). Then, by (a), we have     x − y ≤ x − λz − (1 − λ)y = (x − y) − λ(z − y). Hence, as E is a real inner product space, we get x − y 2 ≤ x − y 2 − 2λx − y, z − y + λ2 z − y 2 , and consequently, λ x − y, z − y ≤ z − y 2 . 2 Thus, (b) follows by letting λ → 0. Conversely, if x ∈ E and y ∈ S satisfy (b), then for every z ∈ S, we have x − y 2 − x − z 2 = 2x − y, z − y − z − y 2 ≤ 0. Thus, x and y satisfy (a).




If E = R2 and S is a closed convex subset of R2 , then condition (b) has a clear geometrical meaning: The angle between the line through x and y and the line through z and y is always obtuse (see Figure 3.3).

Theorem 3.6.6. If H1 is a closed subspace of a Hilbert space H, then every element x ∈ H has a unique decomposition in the form x = y + z where y ∈ H1 and z ∈ H1⊥ . Proof: If x ∈ H1 , then the obvious decomposition is x = x + 0. Suppose

now that x ∈ / H1 . Let y be the unique point of H1 satisfying x − y = infw∈H1 x − w , as in Theorem 3.6.4. We will show that x = y + (x − y) is the desired decomposition. If w ∈ H1 and λ ∈ C, then y + λw ∈ H1 and x − y 2 ≤ x − y − λw 2 = x − y 2 − 2 Re λw, x − y + |λ|2 w 2 . Hence, −2 Re λw, x − y + |λ|2 w 2 ≥ 0.

131

3.6 Orthogonal Complements and Projections

Figure 3.4 Orthogonal decomposition in R2 . If λ > 0, then dividing by λ and letting λ → 0 gives Rew, x − y ≤ 0.

(3.43)

Similarly, replacing λ by −iλ (λ > 0), dividing by λ, and letting λ → 0 gives Imw, x − y ≤ 0.

(3.44)

Since y ∈ H1 implies −y ∈ H1 , inequalities (3.43) and (3.44) hold also with −w instead of w. Therefore, w, x − y = 0 for every w ∈ H1 , which means x − y ∈ H1⊥ . To prove the uniqueness note that if x = y1 + z1 , y1 ∈ H1 , and z1 ∈ H1⊥ , then y − y1 ∈ H1 and z − z1 ∈ H1⊥ . Since y − y1 = z1 − z, we must have y − y1 = z1 − z = 0.
The above theorem can be also stated as follows: If H1 is a closed subspace of a Hilbert space H, then H is the direct sum of H1 and H1⊥ , that is, H = H1 ⊕ H1⊥ .

(3.45)

The representation of H as H1 ⊕H1⊥ is called an orthogonal decomposition of H. Note that the union of a basis of H1 and a basis of H1⊥ is a basis of H. Theorem 3.6.6 allows us to define a mapping PH1 (x) = y, where y ∈ H1 is the unique element such that x = y + z and z ∈ H1 ⊥ . Mapping PH1 is called the orthogonal projection onto H1 . Such mappings will be discussed in Section 4.7.

Example 3.6.7. Let H = R2 . Figure 3.4 exhibits the geometric meaning of the

orthogonal decomposition in R2 . Here x ∈ R2 , x = y + z, y ∈ S, and z ∈ S⊥ . Note that if s0 is a unit vector in S, then y = x, s0 s0 .


Example 3.6.8. If H = R3 , given a plane P, any vector x can be projected onto the plane P. Figure 3.5 illustrates this situation.

Theorem 3.6.9. If S is a closed subspace of a Hilbert space H, then S⊥⊥ = S.




132

Chapter 3

Hilbert Spaces and Orthonormal Systems

Figure 3.5 Orthogonal projection onto a plane.

Proof: If x ∈ S, then for every z ∈ S⊥ , we have x, z = 0, which means

x ∈ S⊥⊥ . Thus, S ⊂ S⊥⊥ . To prove that S⊥⊥ ⊂ S consider an x ∈ S⊥⊥ . Since S is closed, x = y + z for some y ∈ S and z ∈ S⊥ . In view of the inclusion S ⊂ S⊥⊥ , we have y ∈ S⊥⊥ and thus z = x − y ∈ S⊥⊥ , because S⊥⊥ is a vector subspace. But z ∈ S⊥ , so we must have z = 0, which means x = y ∈ S. This shows that S⊥⊥ ⊂ S, completing the proof.


3.7 Linear Functionals and the Riesz Representation Theorem In Section 3.5 we remarked that for any fixed vector x0 in an inner product space E, the formula f (x) = x, x0  defines a bounded linear functional on E. It turns out that if E is a Hilbert space, then every bounded linear functional is of this form. Before proving this result, known as the Riesz representation theorem, we discuss some examples and prove a lemma.

Example 3.7.1. Let H = L2 ((a, b)), −∞ < a < b < ∞. Define a linear functional f on H by the formula 

b

f (x) =

x(t)dt. a

If x0 denotes the constant function 1 on (a, b), then clearly f (x) = x, x0 , and thus f is a bounded functional.


Example 3.7.2. Let H = L2 (a, b), and let t0 be a fixed point in (a, b). Let f be a

functional on H defined by f (x) = x(t0 ). This functional is linear, but it is not bounded.


Example 3.7.3. Let H = Cn , and let n0 ∈ {1, 2, . . . , n}. Define f by the formula   f (x1 , . . . , xn ) = xn0 .

We have

 $ %  f (x1 , . . . , xn ) = (x1 , . . . , xn ), en0 ,

3.7 Riesz Representation Theorem

133

where en0 is the vector which has 1 on the n0 th place and zeros on the remaining places. Thus, f is a bounded linear functional.


Example 3.7.4. Let E be the space of all sequences of complex numbers that have only a finite number of nonzero terms with the inner product defined as in l2 , that is, ∞ $ %  (xn ), (yn ) = xn yn . n

Then f (x) =

∞  xn n

n

defines a bounded linear functional in E. However, there is no (zn ) ∈ E such that f (x) = (xn ), (zn ) for every (xn ) ∈ E. Note that E is not a Hilbert space.


Lemma 3.7.5. If f is a nontrivial bounded linear functional on a Hilbert space H, then dim N (f )⊥ = 1.

Proof: Since f is continuous, N (f ) is a closed proper subspace of H and thus N (f )⊥ is not empty. Let x1 , x2 ∈ N (f )⊥ be nonzero vectors. Since f (x1 ) = 0 and f (x2 ) = 0, there exists a scalar a = 0 such that f (x1 ) + af (x2 ) = f (x1 + ax2 ) = 0. Thus, x1 + ax2 ∈ N (f ). On the other hand, since N (f )⊥ is a vector space and x1 , x2 ∈ N (f )⊥ , we must have x1 + ax2 ∈ N (f )⊥ . This is only possible if x1 + ax2 = 0, which shows that x1 and x2 are linearly dependent, because a = 0.
Example 3.7.6. The property in the above lemma need not hold in incomplete inner product spaces. Consider the space E and the functional f in Example 3.7.4. Then f is clearly a nontrivial bounded linear functional on E, but N (f )⊥ = {0}.


Theorem 3.7.7. (Riesz representation theorem) Let f be a bounded linear functional on a Hilbert space H. There exists exactly one x0 ∈ H such that f (x) = x, x0  for all x ∈ H. Moreover, we have f = x0 .

Proof: If f (x) = 0 for all x ∈ H, then x0 = 0 has the desired properties. Assume now that f is a nontrivial functional. Then dim N (f )⊥ = 1, by Lemma 3.7.5. Let z0 be a unit vector in N (f )⊥ . Then, for every x ∈ H, we have x = x − x, z0 z0 + x, z0 z0 . Since x, z0 z0 ∈ N (f )⊥ , we must have x − x, z0 z0 ∈ N (f ), which means that   f x − x, z0 z0 = 0.

134

Chapter 3

Hilbert Spaces and Orthonormal Systems

Consequently,  $  % f (x) = f x, z0 z0 = x, z0 f (z0 ) = x, f (z0 )z0 . Therefore, if we put x0 = f (z0 )z0 , then f (x) = x, x0  for all x ∈ H. Suppose now that there is another point x1 such that f (x) = x, x1  for all x ∈ H. Then x, x0 − x1  = 0 for all x ∈ H, and thus x0 − x1 , x0 − x1  = 0. This is only possible if x0 = x1 . Finally, we have     f = sup |f (x)| = sup x, x0  ≤ sup x x0 = x0 x =1

and

x =1

x =1

  x0 2 = x0 , x0  = f (x0 ) ≤ f x0 .

Therefore f = x0 .




The collection H of all bounded linear functionals on a Hilbert space H is a Banach space (see Theorem 1.5.9). The Riesz representation theorem says that H = H, or more precisely, H and H are isomorphic. The element x0 corresponding to a functional f is sometimes called the representer of f . Note that the functional f defined by f (x) = x0 , x, where x0 = 0 is a fixed element of a complex Hilbert space H, is not linear. Indeed, we have f (αx + βy) = αf (x) + βf (y). Such functionals are often called antilinear or conjugatelinear. One of the most important theorems in functional analysis is the Hahn– Banach theorem. It says that a bounded linear functional on a subspace of a normed space E can be extended to a bounded linear functional on E. The proof of this fact is not effective. We can only prove that it is possible, but in general we cannot construct such an extension. The situation in Hilbert spaces is much simpler.

Theorem 3.7.8. Let f be bounded linear functional defined on a subspace E of a Hilbert space H. Then there exists a bounded linear functional g defined on H such that (a) f (x) = g(x)

for all x ∈ E,

(b) f = g .

Proof: If E is a closed subspace, then it is a Hilbert space and, by the Riesz representation theorem, there exists an x0 ∈ E such that f (x) = x, x0  for

3.8 Exercises

135

all x ∈ E. But then g can be defined as g(x) = x, x0 . Clearly, (a) and (b) are satisfied. If E is not closed, then we first extend f to a bounded linear functional defined on the closure of E.


3.8 Exercises 1. Show that x, αy + βz = αx, y + βx, z for all α, β ∈ C, in any inner product space. 2. Prove that the space Co (R) of all complex-valued continuous functions that vanish outside some finite interval is an inner product space with the inner product  ∞ f , g = f (x)g(x)dx. −∞

3. Verify that the spaces in Examples 3.2.2–3.2.7 are inner product spaces. 4. (a) Let E = C 1 ([a, b]), the space of all continuously differentiable complex-valued functions on [a, b]. For f , g ∈ E define  f , g =

b

f (x)g (x)dx.

a

Is ·, · an inner product in E ? (b) Let F = {f ∈ C 1 ([a, b]): f (a) = 0}. Is ·, · defined in (a) an inner product in F ? Is it a Hilbert space ? 5. Is the space Co1 (R) of all continuously differentiable complex-valued continuous functions that vanish outside some finite interval an inner product space if  ∞ f , g = f (x)g (x)dx ? −∞

Is it a Hilbert space ? 6. Let E be a normed space, a ∈ E, and r > 0. Show that the ball S(a, r) = {x ∈ E: x − a ≤ r} is a convex set. 7. Show that, in any inner product space, x − y + y − z = x − z if and only if y = αx + (1 − α)z for some α ∈ [0, 1]. Is the same true in a normed space ?

136

Chapter 3

Hilbert Spaces and Orthonormal Systems

8. Let E1 , . . . , En be inner product spaces. Show that % $ [x1 , . . . , xn ], [y1 , . . . , yn ] = x1 , y1  + · · · + xn , yn  defines an inner product in E = E1 × · · · × En . If E1 , . . . , En are Hilbert spaces, show that E is a Hilbert space and its norm is defined by   [x1 , . . . , xn ] = x1 2 + · · · + xn 2 . 9. Show that, in an inner product space, x ⊥ y if and only if x + αy = x − αy for all scalars α. 10. Show that the polarization identity x, y =

 1 x + y 2 − x − y 2 + i x + iy 2 − i x − iy 2 4

holds in any inner product space. 11. Show that, for any x in an inner product space, x = sup y =1 |x, y|. 12. Prove that any complex normed space with norm · satisfying the parallelogram law is an inner product space with the inner product defined by x, y =

 1 x + y 2 − x − y 2 + i x + iy 2 − i x − iy 2 , 4

and then x 2 = x, x. 13. Is C ([a, b]) with the norm f = max[a,b] |f (x)| an inner product space ? 14. Show that L2 ([a, b]) is the only inner product space among the spaces Lp ([a, b]). 15. Show that for any elements in an inner product space,  2   x + y 1 2 2 2 . z− z − x + z − y = x − y + 2  2 2  The equality is called Apollonius’ identity. 16. Prove that any finite dimensional inner product space is a Hilbert space. 17. Complete the proof of completeness of L2,ρ ([a, b]) (Example 3.3.7) by showing that √Fρ ∈ L2,ρ ([a, b]) and fn → √Fρ in L2,ρ ([a, b]). 18. Let E be an incomplete inner product space. Let H be the completion of E (see Section 1.4). Is it possible to extend the inner product from E onto H such that H would become a Hilbert space ?

3.8 Exercises w

137

w

19. Suppose xn → x and yn → y as n → ∞ in a Hilbert space, and αn → α in C. Prove or give a counterexample: w

(a) xn + yn → x + y. w

(b) αn xn → αx. (c) xn , yn  → x, y. (d) xn → x . (e) If xn = yn for all n ∈ N, then x = y. 20. Show that, in a finite dimensional Hilbert space, weak convergence implies strong convergence.  21. If ∞ n=1 un = u, show that ∞  un , x = u, x n=1

for any x in an inner product space. 22. Let {x1 , . . . , xn } be a finite orthonormal set in a Hilbert space H. Prove that for any x ∈ H the vector n  x, xk xk x− k=1

is orthogonal to xk for every k = 1, . . . , n. 23. In the inner product space C ([−π, π ]), show that the following sequences of functions are orthogonal (a) xk (t) = sin kt, k = 1, 2, 3, . . . , (b) yn (t) = cos nt, n = 0, 1, 2, . . . . 24. Show that the application of the Gram–Schmidt process to the sequence of functions f0 (t) = 1, f1 (t) = t, f2 (t) = t 2 , . . . , fn (t) = t n , . . . (as elements of L2 ([−1, 1])) yields the Legendre polynomials. 25. Show that the application of the Gram–Schmidt process to the sequence of functions f0 (t) = e−t f1 (t) = te

2 /2

,

−t 2 /2

,

138

Chapter 3

Hilbert Spaces and Orthonormal Systems

f2 (t) = t 2 e−t fn (t) = t n e

2 /2

−t 2 /2

,..., ,...

(as elements of L2 (R)) yields the orthonormal system discussed in Example 3.4.7. 26. Apply the Gram–Schmidt process to the sequence of functions f0 (t) = 1, f1 (t) = t, f2 (t) = t 2 , . . . , fn (t) = t n , . . . defined on R with the inner product  ∞ 2 f (t)g(t)e−t dt. f , g = −∞

Compare the result with Example 3.4.7. 27. Apply the Gram–Schmidt process to the sequence of functions f0 (t) = 1, f1 (t) = t, f2 (t) = t 2 , . . . , fn (t) = t n , . . . defined on [0, ∞) with the inner product  ∞ f (t)g(t)e−t dt. f , g = 0

The obtained polynomials are called the Laguerre polynomials (Edmond Nicolas Laguerre (1834–1886)). 28. Let Tn be the Chebyshev polynomial (Pafnuty Lvovich Chebyshev (1821– 1894)) of degree n, that is, T0 (x) = 1,

Tn (x) = 21−n cos(n arccos x).

Show that the functions 2n ϕn (x) = √ Tn (x), 2π

n = 0, 1, 2, . . . ,

form an orthonormal system in L2 [(−1, 1)] with respect to the inner product  1 1 f , g = f (x)g(x)dx. √ 1 − x2 −1 29. Prove that for any polynomial pn (x) = xn + an−1 xn−1 + · · · + a0 ,

3.8 Exercises

we have

139

    max pn (x) ≥ max Tn (x),

[−1,1]

[−1,1]

where Tn denotes the Chebyshev polynomial of degree n. 30. Show that the complex functions . ϕn (z) =

n n−1 z , π

n = 1, 2, 3, . . . ,

form an orthonormal system in the space of continuous complex functions defined in the unit disc D = {z ∈ C: |z| ≤ 1} with respect to the inner product  f , g = f (z)g(z)dz. D

31. Prove that the complex functions 1 ψn (z) = √ zn−1 , 2π

n = 1, 2, 3, . . .

form an orthonormal system in the space of continuous complex functions defined on the unit circle C = {z ∈ C: |z| = 1} with respect to the inner product  f , g = f (z)g(z)dz. C

32. With respect to the inner product  f , g =

1 −1

f (x)g(x)ω(x)dx,

where ω(x) = (1 − x)α (1 + x)β and α, β > −1, show that the Jacobi polynomials (Carl Gustav Jacob Jacobi (1804–1851)) Pn(αβ) (x) =

n   (−1)n −α −β d (1 − x) (1 + x) (1 − x)α (1 + x)β (1 − x2 )n n!2n dxn

form an orthogonal system. 33. Show that the Gegenbauer polynomials (Leopold Bernhard Gegenbauer (1849–1903)) Cnγ (x) =

n    (−1)n  2 1/2−γ d 2 n+γ −1/2 , 1 − x 1 − x n!2n dxn

140

Chapter 3

Hilbert Spaces and Orthonormal Systems

where γ > 12 , form an orthonormal system with respect to the inner product  1  1/2−γ f , g = f (x)g(x) 1 − x2 dx. −1

Note that the Gegenbauer polynomials are a special case of the Jacobi polynomials with α = β = γ − 12 . 34. Find a, b, c ∈ C, which minimize the value of the integral 

1 −1

 x3 − a − bx − cx2 2 dx.

35. If x and xk (k = 1, . . . , n) belong to a real Hilbert space, show that 2  n n n  n        αk xk  = x 2 − αk x, xk  + αk αl xk , xl . x −   k=1

k=1

k=1 l=1

Also show that this expression is minimum when Aa = b where a = (α1 , . . . , αn ), b = (x, x1 , . . . , x, xn ) and the matrix A = (αkl ) is defined by αkl = xk , xl . 36. If (an ) is an orthonormal sequence in a Hilbert space H and (αn ) is a sequence in l2 , show that there exists x ∈ H such that   x, an  = αn and (αn ) = x , where (αn ) denotes the norm in l2 . 37. If αn and βn (n = 1, 2, 3, . . .) are generalized Fourier coefficients of vectors x and y with respect to a complete orthonormal sequence in a Hilbert space, show that x, y =

∞ 

αk β k .

k=1

38. Let (xn ) be an orthonormal sequence in a Hilbert space H. Show that (xn ) is complete if and only if cl(span{x1 , x2 , . . .}) = H. In other words, (xn ) is complete if and only if every element of H can be approximated by a sequence of finite combinations of xn ’s. 39. Show that the functions ϕn (x) =

e−x/2 Ln (x), n!

n = 0, 1, 2, . . .

3.8 Exercises

141

where Ln is the Laguerre polynomial of degree n, that is, Ln (x) = ex

dn  n −x  x e , dxn

form a complete orthonormal system in L2 (0, ∞). 40. Let einx ϕn (x) = √ , 2π

n = 0, ±1, ±2, . . . ,

and let f ∈ L1 ([−π, π ]). Define fn (x) =

n 

f , ϕk ϕk ,

for n = 0, 1, 2, . . . .

k=−n

Show that

 n   |k| f0 (x) + f1 (x) + · · · + fn (x) = f , ϕk ϕk (x). 1− n+1 n+1 k=−n

41. Fill in the details of the proof of Lemma 3.5.1. 42. Let f be a continuous non-negative function defined on [−π,  π π ] such that supp f ⊆ [−π + ε, π − ε], for some 0 < ε < π , and −π f (x)dx = 2π . Let g be a 2π -periodic extension of f onto the entire line R. Define kn (x) = ng(nx)

for n = 1, 2, . . . .

Show that (kn ) is a summability kernel. 43. Show that the sequence of functions 1 cos x sin x cos 2x sin 2x √ , √ , √ , √ , √ ,..., π π π π 2π is a complete orthonormal sequence in L2 ([−π, π ]). 44. Show that the following sequence of functions is a complete orthonormal system in L2 ([0, π ]): . . . 1 2 2 2 cos x, cos 2x, cos 3x, . . . . √ , π π π π 45. Show that the following sequence of functions is a complete orthonormal system in L2 ([0, π ]): . . . 2 2 2 sin x, sin 2x, sin 3x, . . . . π π π

142

Chapter 3

Hilbert Spaces and Orthonormal Systems

46. Give an example of a complete orthonormal sequence in L2 ([a, b]) for arbitrary a < b. 47. What is the orthogonal complement in L2 (R) of the set of all even functions ? 48. What is the orthogonal complement in L2 ([−π, π ]) of the set of all polynomials of odd degree ? 49. Let P be a complete orthonormal system in a Hilbert space H. Show that if P = P1 ∪ P2 and P1 ∩ P2 = ∅, then P1⊥ = cl(span P2 ). 50. Let S be a subset of an inner product space. Show that S⊥ = (span S)⊥ . 51. Prove the following generalization of Theorem 3.6.5 for complex inner product spaces: Let S be a convex subset of a complex inner product space E, y ∈ S and let x ∈ E. Then the following conditions are equivalent: (a) x − y = infz∈S x − z , (b) Rex − y, z − y ≤ 0 for all z ∈ S. 52. Let E be the Banach space R2 with the norm (x, y) = max{|x|, |y|}. Show that E does not have the closest point property. 53. Let S be a closed subspace of a Hilbert space H and let (e1 , e2 , . . .) be a complete orthonormal sequence in S. For an arbitrary x ∈ H there exists y ∈ S such that x − y = infz∈S x − z . Define y in terms of (e1 , e2 , . . .). 54. If S is a closed subspace of a Hilbert space H, then H = S ⊕ S⊥ . Is this true in every inner product space ? 55. Show that the functional in Example 3.7.2 is unbounded. 56. The Riesz representation theorem says, that for every bounded linear functional f ∈ H on a Hilbert space H, there exists a representer xf ∈ H such that f (x) = x, xf  for all x ∈ H. Let T : H → H be the mapping which assigns xf to f . Prove the following properties of T: (a) T(H ) = H, (b) T(f + g) = T(f ) + T(g), (c) T(αf ) = αT(f ), (d) T(f ) = f , where f , g ∈ H and α ∈ C. 57. Prove part (b) of Theorem 3.4.27. 58. Show that the space l2 is separable.

3.8 Exercises

143

59. Let P be an uncountable orthonormal system in an inner product space E. Show that, for every x ∈ E, we have x, e = 0 for at most countably many e ∈ P .

This page intentionally left blank

Linear Operators on Hilbert Spaces

CHAPTER

4

“As long as a branch of science affords an abundance of problems, it is full of life; want of problems means death or cessation of independent development. Just as every human enterprise prosecutes final aims, so mathematical research needs problems. Their solution steels the force of the investigator; thus he discovers new methods and viewpoints and widens his horizon.” David Hilbert

4.1 Introduction In Section 1.5, we discussed some basic properties of linear mappings between vector spaces. In this chapter, we are interested in the special case when the domain and the range are subspaces of a Hilbert space. In this case, the name linear operator or linear transformation is usually used. Since nonlinear operators will not be considered here, linear operators will be called simply operators. Operators on Hilbert spaces are widely used to represent physical quantities, and hence their importance is further enhanced in applied mathematics and mathematical physics. The most important operators include differential, integral, and matrix operators. In the next section, we present some examples of operators. Then bilinear functionals and quadratic forms are considered and the Lax–Milgram theorem is proved. This theorem is an important generalization of the Riesz representation theorem. Adjoint and self-adjoint operators are discussed in Section 4.4. Sections 4.5 to 4.8 deal with special linear operators including invertible, normal, isometric, unitary, positive, compact, and projection operators. In Section 4.9, we consider eigenvalues and eigenvectors of linear operators. These concepts play a central role in the theory of operators

145

146

Chapter 4

Linear Operators on Hilbert Spaces

and their applications, where the spectral decomposition of operators is one of the most important tools. The spectral theorem for compact self-adjoint operators and other related results are discussed in Section 4.10. Basic concepts of the theory of unbounded operators are discussed in Section 4.12.

4.2 Examples of Operators We begin with some examples of operators. In each case, we are interested whether the operator is bounded. Recall that an operator A is called bounded if there is a number K such that Ax ≤ K x for every x in the domain of A. The norm of A is defined as the infimum of all such numbers K, or equivalently, by A = sup Ax . x =1

We will refer to this norm as the operator norm. In Section 1.6, we proved that an operator A is bounded if and only if it is continuous. It is often much more difficult to find the norm of an operator than just to prove that it is bounded.

Example 4.2.1. (Identity operator and null operator) The simplest examples

of operators are the identity operator I and the null operator. The identity operator leaves every element unchanged, that is, I x = x for all x ∈ E. The null operator assigns the zero vector to every element of E. The null operator will be denoted by 0. Obviously, the identity operator and the null operator are bounded and we have I = 1 and 0 = 0. A scalar multiple α I of the identity operator is the operator, which multiplies every element by the scalar α, that is, (α I )x = αx.


Example 4.2.2. Let A be an operator on CN and let {e1 , . . . , eN } be the stan-

dard orthonormal base in CN , that is,

e1 = (1, 0, 0, . . . , 0), e2 = (0, 1, 0, . . . , 0), .. . eN = (0, 0, . . . , 0, 1). Define, for i, j ∈ {1, 2, . . . , N}, αij = Aej , ei .

4.2 Examples of Operators

Then, for x =

N

j=1 λj ej

∈ CN , we have Ax =

Ax, ei  =

N 

N

λj Aej , ei  =

j=1

j=1 λj Aej , N 

147

and hence

αij λj .

(4.1)

j=1

Thus, every operator on the space CN is defined by an N × N matrix. Conversely, for every N × N matrix (αij ), formula (4.1) defines an operator on CN . We thus have a one-to-one correspondence between operators on Ndimensional vector spaces and N × N matrices. If operator A is defined by the matrix (αij ), then   N N   A ≤  |αij |2 . i=1 j=1

This implies that every operator on CN , and thus every operator on any finite dimensional Hilbert space, is bounded.
The operator in the next example is defined on a proper subspace of a Hilbert space.

Example 4.2.3. (Differential operator) One of the most important operators in applied mathematics is the differential operator (Df )(x) =

df (x) = f (x) dx

defined on a space of differentiable functions. Consider, for example, the differential operator on a subspace of L2 ([−π, π ]) defined as
    D(D) = f ∈ L2 [−π, π ] : f ∈ L2 [−π, π ] .

 π 2 If L2 ([−π, π ]) is equipped with the standard norm f = −π |f (x)| dx, then the differential operator is not bounded. Indeed, for fn (x) = sin nx, n = 1, 2, 3, . . . , we have + π √ |sin nx|2 dx = π fn = −π

and

+ Dfn =

π −π

√ |n cos nx|2 dx = n π.

This example can be easily generalized to an arbitrary interval [a, b] or even (−∞, ∞).


148

Chapter 4

Linear Operators on Hilbert Spaces

Example 4.2.4. (Integral operator) Another important type of operators is an integral operator T defined by 

b

(Tx)(s) =

K(s, t)x(t)dt, a

where a and b are finite or infinite, a < b, and K is a function defined on the square (a, b) × (a, b). The function K is called the kernel of the operator. The domain of an integral operator depends on K. If  b a

b

 K(s, t)2 dtds < ∞,

a

then T is a bounded operator on L2 ([a, b]) and +  b b   K(s, t)2 dtds. T ≤ a

a

Indeed, for any x ∈ L2 ([a, b]), we have 2  b  b    ds  Tx 2 = K(s, t)x(t)dt   a

≤

a

 b  a

≤

 K(s, t)2 dt

  x(t)2 dt ds



a

 b a

b

b

(by Schwarz’s inequality)

a

b

 K(s, t)2 dtds



a

b

 x(t)2 dt.

a

Thus, + Tx ≤

 b a

b

 K(s, t)2 dtds x .




a

Example 4.2.5. (Multiplication operator) Let z ∈ C ([a, b]). An operator A

on L2 ([a, b]) defined by (Ax)(t) = z(t)x(t) is clearly linear. The function z is called the multiplier. Since 

b

    x(t)2 z(t)2 dt ≤ maxz(t)2

Ax = 2

[a,b]

a

we have



b

 x(t)2 dt,

a

  Ax ≤ maxz(t) x , [a,b]

and thus A is bounded.




4.2 Examples of Operators

149

Two operators A and B are said to be equal, A = B, if D (A) = D (B) and Ax = Bx for every x in that common domain. The sum of two operators is defined by (A + B)x = Ax + Bx and D (A + B) = D (A) ∩ D (B). Since we are always assuming that the domain of an operator is a vector subspace, D (A) ∩ D (B) is never empty because 0 ∈ D(A) ∩ D(B). However, in the extreme case, we might have D(A) ∩ D(B) = {0}. The product of a scalar λ and an operator A is defined by (λA)x = λ(Ax). Obviously, D (λA) = D (A). The product AB of operators A and B is defined by (AB)x = A(Bx). The domain of the product of A and B is
 D(AB) = x ∈ D(B): Bx ∈ D(A) . The product AB is simply the composition of A and B. Traditionally, in the context of operators, the word product is used instead of composition. When an operator A is multiplied by a scalar λ, then the result can also be viewed as the product of the multiplication operator λI and A, that is, λA = (I λ)A. This different interpretation does not change the properties of the operation. In some situations, it may be convenient to identify scalars with multiplication operators. Since I A = 1A, the identity operator is often denoted by 1. The defined operations have the following obvious properties: A + B = B + A, α(A + B) = αA + αB, A(BC) = (AB)C,

(A + B) + C = A + (B + C), (α + β)A = αA + βA, (A + B)C = AC + BC,

A + 0 = A, A0 = 0, AI = I A.

It is somewhat unexpected that the equality A(B + C) = AB + AC need not hold in general. The product of operators is not commutative, that is, AB need not equal BA. Operators A and B for which AB = BA are called commuting operators.

Example 4.2.6. (Noncommuting operators) Let A and B be operators on C2 defined by the matrices  A=

1 0

0 0



(see Example 4.2.2). Then AB = BA.

 and

B=

0 1 0 0






150

Chapter 4

Linear Operators on Hilbert Spaces

Example 4.2.7. (Noncommuting operators) Consider the operators Af (x) = xf (x)

D=

and

d . dx

It is easy to check that AD = DA.




The square of an operator A is defined as A2 x = A(Ax). By induction, we can define   An x = A An−1 x for any positive integer n. As usual, A1 = A and A0 = I . If A and B are bounded operators, then obviously A+B and λA are bounded (for any scalar λ) and we have A + B ≤ A + B

and

λA = |λ| A .

A similar property holds for the product of operators.

Theorem 4.2.8. The product AB of bounded operators A and B is bounded and AB ≤ A B .

Proof: Let A and B be bounded operators on a normed space E, A = K1 and B = K2 . Then

ABx ≤ K1 Bx ≤ K1 K2 x , for every x ∈ E.




Theorem 4.2.9. A bounded operator on a separable infinite dimensional Hilbert space can be represented by an infinite matrix. Proof: Let A be a bounded operator on a Hilbert space H and let (en ), n = 1, 2, 3, . . . , be a complete orthonormal sequence in H. For i, j ∈ N, define αij = Aej , ei . For any x ∈ H, we have ∞    n   Ax = A x, ej ej = A lim x, ej ej j=1

= lim A n→∞

n→∞



n  x, ej ej j=1



j=1

(by continuity of A)

4.3 Bilinear Functionals and Quadratic Forms  n   (by linearity of A) = lim x, ej Aej n→∞

=

∞ 

151

j=1

x, ej Aej .

j=1

Now

) ∞ ∞ ∞    Ax, ei  = x, ej Aej , ei = Aej , ei x, ej  = αij x, ej . (

j=1

j=1

j=1

Thus, A is represented by the matrix (aij ).




4.3 Bilinear Functionals and Quadratic Forms The concepts of a bilinear functional and a quadratic form do not require the structure of an inner product space. They can be defined in any vector space. This discussion is presented in this chapter because of important applications to the theory of operators in Hilbert spaces.

Definition 4.3.1. (Bilinear functional) By a bilinear functional ϕ on a complex vector space E, we mean a mapping ϕ : E × E → C satisfying the following two conditions: (a) ϕ(αx1 + βx2 , y) = αϕ(x1 , y) + βϕ(x2 , y), (b) ϕ(x, αy1 + βy2 ) = αϕ(x, y1 ) + βϕ(x, y2 ), for any scalars α and β and any x, x1 , x2 , y, y1 , y2 ∈ E. Bilinear functionals are often called sesquilinear. Note that a bilinear functional is linear with respect to the first variable and antilinear with respect to the second variable. Clearly, all bilinear functionals on E constitute a vector space.

Example 4.3.2. Inner product is a bilinear functional.




Example 4.3.3. Let A and B be operators on an inner product space E. Then ϕ1 (x, y) = Ax, y, ϕ2 (x, y) = x, By, and ϕ3 (x, y) = Ax, By are bilinear functionals on E.


Example 4.3.4. Let f and g be linear functionals on a vector space E. Then ϕ(x, y) = f (x)g(y) is a bilinear functional on E.




152

Chapter 4

Definition 4.3.5.

Linear Operators on Hilbert Spaces

(Symmetric, positive, strictly positive, and bounded bilinear

functionals) Let ϕ be a bilinear functional on E. (a) ϕ is called symmetric if ϕ(x, y) = ϕ(y, x) for all x, y ∈ E. (b) ϕ is called positive if ϕ(x, x) ≥ 0 for every x ∈ E. (c) ϕ is called strictly positive if it is positive and ϕ(x, x) > 0 for all x = 0. (d) If E is a normed space, then ϕ is called bounded if |ϕ(x, y)| ≤ K x y for some K > 0 and all x, y ∈ E. The norm of a bounded bilinear functional is defined by   ϕ = sup ϕ(x, y). x = y =1

If f = g in Example 4.3.4, then ϕ is symmetric and positive. Inner product is strictly positive. If operators A and B in Example 4.3.3 are bounded, then ϕ1 , ϕ2 , and ϕ3 are bounded. Similarly, if f and g in Example 4.3.4 are bounded, then the defined bilinear functional is also bounded. Note that for a bounded bilinear functional ϕ on E we have   ϕ(x, y) ≤ ϕ x y for all x, y ∈ E.

Definition 4.3.6.

(Quadratic form) Let ϕ be a bilinear functional on a vector space E. The function  : E → C defined by (x) = ϕ(x, x) is called the quadratic form associated with ϕ . A quadratic form  on a normed space E is called bounded if there exists a constant K > 0 such that |(x)| ≤ K x 2 for all x ∈ E. The norm of a bounded quadratic form is defined by    = sup (x). x =1

Note that for a bounded quadratic form  on a normed space we have |(x)| ≤  x 2 . A bilinear functional and the associated quadratic form have properties similar to an inner product x, y and the square of the norm defined by that inner product x 2 = x, x, respectively.

Theorem 4.3.7. (Polarization identity) Let ϕ be a bilinear functional on E, and let  be the quadratic form associated with ϕ. Then 4ϕ(x, y) = (x + y) − (x − y) + i(x + iy) − i(x − iy)

(4.2)

4.3 Bilinear Functionals and Quadratic Forms

153

for all x, y ∈ E.

Proof: For any α, β ∈ C, we have (αx + βy) = ϕ(αx + βy, αx + βy) = |α|2 (x) + αβϕ(x, y) + αβϕ(y, x) + |β|2 (y). Using this equality subsequently for α = β = 1; α = 1 and β = −1; α = 1 and β = i; α = 1 and β = −i; we get (x + y) = (x) + ϕ(x, y) + ϕ(y, x) + (y), −(x − y) = −(x) + ϕ(x, y) + ϕ(y, x) − (y), i(x + iy) = i(x) + ϕ(x, y) − ϕ(y, x) + i(y), −i(x − iy) = −i(x) + ϕ(x, y) − ϕ(y, x) − i(y). By adding these equalities we obtain (4.2).




The following simple, but somewhat surprising, result is often useful.

Corollary 4.3.8. Let ϕ1 and ϕ2 be bilinear functionals on E. If ϕ1 (x, x) = ϕ2 (x, x) for all x ∈ E, then ϕ1 = ϕ2 , that is, ϕ1 (x, y) = ϕ2 (x, y) for all x, y ∈ E. Similarly, if A and B are operators on E such that Ax, x = Bx, x for all x ∈ E, then A = B. Proof: If ϕ1 (x, x) = ϕ2 (x, x) for all x ∈ E, then the quadratic forms 1 and 2 associated with ϕ1 and ϕ2 , respectively, are equal, and hence, by (4.2), the functionals ϕ1 and ϕ2 are equal. The proof for operators is obtained by letting ϕ1 (x, y) = Ax, y and ϕ2 (x, y) = Bx, y.
Theorem 4.3.9. A bilinear functional ϕ on E is symmetric if and only if the associated quadratic form  is real. Proof: If ϕ(x, y) = ϕ(y, x) for all x, y ∈ E, then (x) = ϕ(x, x) = ϕ(x, x) = (x) for every x ∈ E, and thus  is real. Assume now (x) = (x) for all x ∈ E. Define a bilinear functional ψ on E by ψ(x, y) = ϕ(y, x). Then, for the associated quadratic form  we have (x) = ϕ(x, x) = (x) = (x). Thus, by Corollary 4.3.8, ϕ(x, y) = ψ(x, y) for all x, y ∈ E. Clearly, this means that ϕ(x, y) = ϕ(y, x) for all x, y ∈ E.


154

Chapter 4

Linear Operators on Hilbert Spaces

Theorem 4.3.10. A bilinear functional ϕ on a normed space E is bounded if and only if the associated quadratic form  is bounded. Moreover, we have  ≤ ϕ ≤ 2  .

(4.3)

Proof: Since      = sup (x) = sup ϕ(x, x) ≤ x =1

x =1

sup

  ϕ(x, y) = ϕ ,

x = y =1

if ϕ is bounded, then  is bounded and the first inequality follows. Suppose now that  is bounded. In view of (4.2), we have     ϕ(x, y) = 1 (x + y) − (x − y) + i(x + iy) − i(x − iy) 4   1 ≤  x + y 2 + x − y 2 + x + iy 2 + x − iy 2 . 4 Hence, by the parallelogram law,     ϕ(x, y) ≤  x 2 + y 2 . Consequently, sup

  ϕ(x, y) ≤

x = y =1

sup

x = y =1

   x 2 + y 2 = 2  .

Thus, if  is bounded, then ϕ is bounded and the second inequality in (4.3) follows.


Theorem 4.3.11. Let ϕ be a bilinear functional on a normed space E and let  be the associated quadratic form. If ϕ is symmetric and bounded, then ϕ =  .

Proof: By Theorem 4.3.10,  ≤ ϕ . We need to show that the opposite inequality holds as well. Since ϕ is symmetric,  is real, by Theorem 4.3.9. Then, by the polarization identity, we obtain Re ϕ(x, y) =

 1 (x + y) − (x − y) , 4

and hence     Re ϕ(x, y) ≤ 1  x + y 2 + x − y 2 4   1 =  x 2 + y 2 , 2 by the parallelogram law. Let x and y be arbitrary fixed elements of E such that x = y = 1, and let θ be a complex number such that |θ | = 1 and

4.3 Bilinear Functionals and Quadratic Forms

155

|ϕ(x, y)| = θ ϕ(x, y). Then     ϕ(x, y) = θ ϕ(x, y) = ϕ(θ x, y) = Re ϕ(θ x, y)   1 ≤  θ x 2 + y 2 =  , 2 and thus ϕ =

sup

  ϕ(x, y) ≤  .

x = y =1




Theorem 4.3.12. Let A be a bounded operator on a Hilbert space H. Then the bilinear functional defined by ϕ(x, y) = Ax, y is bounded and A = ϕ . Proof: For all x, y ∈ H, by Schwarz’s inequality, we have     ϕ(x, y) = Ax, y ≤ Ax y ≤ A x y . Thus, ϕ is bounded and ϕ ≤ A . On the other hand, we have     Ax 2 = Ax, Ax = ϕx, Ax ≤ ϕ Ax Ax . Therefore, for Ax = 0, we have Ax ≤ ϕ x . Since the above inequality is trivially satisfied if Ax = 0, we obtain A ≤ ϕ .
It turns out that every bounded bilinear functional on a Hilbert space is of the form considered in the above theorem. Note that the theorem remains true if ϕ(x, y) = Ax, y is replaced by ϕ(x, y) = x, Ay. The same is true about the next theorem. Our reason for choosing Ax, y in Theorem 4.3.12 and x, Ay in the theorem that follows is that it is more convenient when adjoint operators are introduced.

Theorem 4.3.13. Let ϕ be a bounded bilinear functional on a Hilbert space H. There exists a unique bounded operator A on H such that ϕ(x, y) = x, Ay for all x, y ∈ H.

Proof: For a fixed y ∈ H, ϕ(x, y) is a bounded linear functional on H. Thus, by the Riesz representation theorem, there exists a unique element Ay ∈ H such that ϕ(x, y) = x, Ay for all x ∈ H. We have to prove that the mapping y → Ay is a bounded operator on E. Indeed, for any x, y1 , y2 ∈ H and α, β ∈ C we have

156

Chapter 4

Linear Operators on Hilbert Spaces

% $ x, A(αy1 + βy2 ) = ϕ(x, αy1 + βy2 ) = αϕ(x, y1 ) + βϕ(x, y2 ) = x, αAy1 + βAy2 , and thus A(αy1 + βy2 ) = αAy1 + βAy2 . Now we show that A is bounded. Since ϕ is bounded, we have     x, Ay = ϕ(x, y) ≤ k x y for some k > 0 and all x, y ∈ H. In particular, for x = Ay we obtain     Ay 2 = Ay, Ay = ϕ(Ay, y) ≤ k Ay y . Therefore, if Ay = 0, we get Ay ≤ k y , which is also trivially satisfied if Ay = 0. This proves that A is bounded. To prove uniqueness, notice that x, Ay = x, By

for all x, y ∈ H

implies A = B.




Definition 4.3.14.

(Coercive functional) A bilinear functional ϕ on a normed space E is called coercive (or elliptic) if there exists a positive constant K such that ϕ(x, x) ≥ K x 2

for all x ∈ E.

Example 4.3.15. If z is a continuous real-valued function on [0, 1] such that mint∈[0,1] z(t) > 0, then the bilinear functional ϕ defined on L2 ([0, 1]) by  1 x(t)y(t)z(t)dt ϕ(x, y) = 0

is coercive. Indeed, we have 

1

 x(t)2 z(t)dt ≥ K x 2

ϕ(x, x) = 0

where K = mint∈[0,1] z(t).




4.3 Bilinear Functionals and Quadratic Forms

157

The following theorem, proved by P. Lax and A.N. Milgram in 1954, is an important generalization of the Riesz representation theorem. Applications of this theorem to boundary value problems are discussed in Chapter 6.

Theorem 4.3.16. (Lax–Milgram theorem) Let ϕ be a bounded, coercive, bilinear functional on a Hilbert space H. For every bounded linear functional f on H, there exists a unique xf ∈ H such that f (x) = ϕ(x, xf )

for all x ∈ H.

Proof: By Theorem 4.3.13, there exists a bounded operator A such that ϕ(x, y) = x, Ay

for all x, y ∈ H.

Since ϕ is coercive, we have K x 2 ≤ ϕ(x, x) = x, Ax ≤ Ax x , and hence K x ≤ Ax

for all x ∈ H.

Let x1 , x2 ∈ H. If Ax1 = Ax2 , then A(x1 − x2 ) = 0, and thus x1 − x2 ≤

 1 A(x1 − x2 ) = 0, K

which implies x1 = x2 . Therefore, A is one-to-one. Denote the range of A by R(A). Let (xn ) be a sequence of elements of H. If Axn − y → 0, for some y ∈ H, then xn − xm ≤

1 Axn − Axm → 0, K

as m, n → ∞. Therefore (xn ) is a Cauchy sequence in H. Since H is complete, there exists x ∈ H such that xn − x → 0. Hence Axn − Ax → 0, because A is continuous. Consequently, Ax = y, and thus y ∈ R(A). This proves that R(A) is a closed subspace of H. We will prove that R(A) = H. Suppose R(A) is a proper subspace of H. Then, by Theorem 3.6.6, there exists a nonzero x ∈ H, which is orthogonal to R(A), that is, x, Ay = 0 for all y ∈ H. In particular, we have     0 = x, Ax = ϕ(x, x) ≥ K x 2 , which contradicts the assumption x = 0.

158

Chapter 4

Linear Operators on Hilbert Spaces

If f is a bounded linear functional on H, then there exists a unique x0 ∈ H such that f (x) = x, x0  for all x ∈ H. Since A is a one-to-one mapping and R(A) = H, there exists a unique xf ∈ H such that x0 = Axf , and hence f (x) = x, Axf  = ϕ(x, xf )

for all x ∈ H.




4.4 Adjoint and Self-Adjoint Operators Consider a bounded operator A on a Hilbert space H. Since the bilinear functional ϕ(x, y) = Ax, y is bounded (Theorem 4.3.12), there exists a unique bounded operator A∗ such that Ax, y = ϕ(x, y) = x, A∗ y for all x, y ∈ H (Theorem 4.3.13).

Definition 4.4.1.

(Adjoint operator) Let A be a bounded operator on a Hilbert space H . The operator A∗ : H → H defined by $ % Ax, y = x, A∗ y for all x, y ∈ H is called the adjoint operator of A. Adjoint operators, as just defined, are sometimes called Hilbert-adjoint to distinguish them from the more general definition considered later in this chapter. The following properties are direct consequences of Definition 4.4.1: (A + B)∗ = A∗ + B∗ , (αA)∗ = αA∗ ,  ∗ ∗ A = A,

I∗ = I, (AB)∗ = B∗ A∗ ; for arbitrary operators A and B and scalar α.

Theorem 4.4.2. The adjoint operator A∗ of a bounded operator A is bounded.

Moreover, we have A = A∗ and A∗ A = A 2 .

Proof: The argument presented before Definition 4.4.1 shows that the adjoint operator A∗ of a bounded operator A is bounded. Since Ax, y

4.4 Adjoint and Self-Adjoint Operators

159

and x, A∗ y define the same bilinear functional, Theorem 4.3.12 gives us A = A∗ . It follows from Theorem 4.2.8 that  ∗   ∗ A A ≤ A  A = A 2 . On the other hand, for every x ∈ H, we have    $ %  Ax 2 = Ax, Ax = A∗ Ax, x ≤ A∗ Ax x ≤ A∗ A x 2 . Therefore A∗ A = A 2 .




Operators A and A∗ need not be equal. For instance, let H = C2 and let A be defined by A(z1 , z2 ) = (0, z1 ). Then $ % A(x1 , x2 ), (y1 , y2 ) = x1 y2

and

$ % (x1 , x2 ), A(y1 , y2 ) = x2 y1 .

Operators for which A = A∗ are of special interest.

Definition 4.4.3.

(Self-adjoint operator) If A = A , then A is called self-adjoint. ∗

In other words, if A is self-adjoint, then Ax, y = x, Ay for all x, y ∈ H.

Example 4.4.4. Let H = CN and let {e1 , . . . , eN } be the standard orthonormal base in H. Let A be an operator represented by matrix (aij ), where aij = Aej , ei , (see Example 4.2.2). Then the adjoint operator A∗ is represented by the matrix bkj = A∗ ej , ek . Consequently bkj = ej , Aek  = Aek , ej  = ajk . Therefore, the operator A is self-adjoint if and only if aij = aji . A matrix satisfying this condition is often called Hermitian.


Example 4.4.5. Let H be a separable, infinite dimensional Hilbert space and let {e1 , e2 , e3 , . . .} be a complete orthonormal sequence in H. Let A be a bounded operator on H represented by an infinite matrix (aij ) (see Theorem 4.2.9). As in the finite dimensional case, the adjoint operator A∗ is represented by the infinite matrix (aji ). A is self-adjoint if and only if aij = aji for all i, j ∈ N.


160

Chapter 4

Linear Operators on Hilbert Spaces

Example 4.4.6. Let T be a Fredholm operator on L2 ([a, b]) defined by 

b

(Tx)(s) =

K(s, t)x(t)dt, a

where K is a function defined on [a, b] × [a, b] such that  b a

b

 K(s, t)2 dsdt < ∞.

a

Note that the condition is satisfied if K is continuous. We have  b b K(s, t)x(t)y(s)dsdt Tx, y = a

=

a

 b

b

K(s, t)x(t)y(s)dsdt a



a



b

=

K(s, t)y(s)dsdt.

a

This shows that 

 T x (s) = ∗

b

x(t) a



b

K(t, s)x(t)dt. a

Thus, a Fredholm operator is self-adjoint if its kernel satisfies the equality K(s, t) = K(t, s).


Example 4.4.7. Let A be the operator on L2 ([a, b]) defined by (Ax)(t) = tx(t). Since



b

Ax, y = a



b

tx(t)y(t)dt =

x(t)ty(t)dt = x, Ay,

a

A is self-adjoint.




Example 4.4.8. Consider the operator A defined on L2 (R) defined by (Ax)(t) = e−|t| x(t). This is a bounded self-adjoint operator. Boundedness of A can be shown as in Example 4.2.5. Moreover, we have  ∞  ∞   −|t| Ax, y = e x(t)y(t)dt = x(t) e−|t| y(t) dt = x, Ay. −∞

Thus, A is self-adjoint.

−∞




4.4 Adjoint and Self-Adjoint Operators

161

Now we prove some theorems that can be used to construct more examples of self-adjoint operators.

Theorem 4.4.9. Let ϕ be a bounded bilinear functional on H and let A be an operator on H such that ϕ(x, y) = x, Ay for all x, y ∈ H. Then A is self-adjoint if and only if ϕ is symmetric.

Proof: For all x, y ∈ H, we have x, Ay = ϕ(x, y) = ϕ(y, x) = y, Ax = Ax, y

(if ϕ is symmetric)

and ϕ(x, y) = x, Ay = Ax, y = y, Ax = ϕ(y, x)

(if A is self-adjoint).




Theorem 4.4.10. Let A be a bounded operator on a Hilbert space H. The operators T1 = A∗ A and T2 = A + A∗ are self-adjoint. Proof: For all x, y ∈ H, we have $ % $ % T1 x, y = A∗ Ax, y = Ax, Ay = x, A∗ Ay = x, T1 y and $  % $  ∗ % $   % T2 x, y = A + A∗ x, y = x, A + A∗ y = x, A + A∗ y = x, T2 y.


Theorem 4.4.11. The product of two self-adjoint operators is self-adjoint if and only if the operators commute.

Proof: Let A and B be self-adjoint operators. Then ABx, y = Bx, Ay = x, BAy. Thus, if AB = BA, then AB is self-adjoint. Conversely, if AB is self-adjoint, then the above implies AB = (AB)∗ = BA.


Corollary 4.4.12. If A is self-adjoint, then so is any polynomial of A with real coefficients αn , . . . , α0 : α n An + · · · + α 1 A + α0 I .

Theorem 4.4.13. For every bounded operator T on a Hilbert space H, there exist unique self-adjoint operators A and B such that T = A + iB and T ∗ = A − iB. Proof: Let T be a bounded operator on H. Define A=

 1 T + T∗ 2

and

B=

 1 T − T∗ . 2i

162

Chapter 4

Linear Operators on Hilbert Spaces

Clearly, A and B are self-adjoint, and T = A + iB. Moreover, for any x, y ∈ H, we have $ % Tx, y = (A + iB)x, y = Ax, y + iBx, y = x, Ay + ix, By $ % = x, (A − iB)y . Hence, T ∗ = A − iB. Proving the uniqueness is left as an exercise.




In particular, if T is self-adjoint, then A = T and B = 0. Self-adjoint operators are like real numbers in C. The following property of self-adjoint operators will be useful when we discuss spectral properties of such operators in Sections 4.9 and 4.10.

Theorem 4.4.14. Let T be a self-adjoint operator on a Hilbert space H. Then   T = sup Tx, x.

(4.4)

x =1

Proof: Let

  M = sup Tx, x. x =1

If x = 1, then   Tx, x ≤ Tx x = Tx ≤ T x = T . Thus, M ≤ T .

(4.5)

On the other hand, for all x, z ∈ H, we have $ % $ %   T(x + z), x + z − T(x − z), x − z = 2 Tx, z + Tz, x = 4 ReTx, z. Therefore, ReTx, z ≤

 M 2  M x + z 2 + x − z 2 = x + z 2 . 4 2

Now suppose x = 1 and Tx = 0. If we let z =

Tx Tx ,

then

2 1 Tx = Tx , ReTx, z = Re Tx, Tx

(4.6)

4.5 Normal, Isometric, and Unitary Operators

163

and, by (4.6), ReTx, z ≤

    Tx 2 M  x 2 +   Tx  = M. 2

Consequently, T ≤ M.




4.5 Invertible, Normal, Isometric, and Unitary Operators Definition 4.5.1. (Inverse operator) Let A be an operator defined on a vector subspace of E. An operator B defined on R(A) is called the inverse of A if ABx = x for all x ∈ R(A) and BAx = x for all x ∈ D (A). An operator which has an inverse is called invertible. The inverse of A is denoted by A−1 . If an operator has an inverse, then it is unique. Indeed, suppose B1 and B2 are inverses of A. Then B1 = B1 I = B1 AB2 = I B2 = B2 . Note also that     D A−1 = R(A) and R A−1 = D(A). In the next theorem, we recall some simple algebraic properties of invertible operators. Easy proofs are left as exercises.

Theorem 4.5.2. (a) The inverse of a linear operator is a linear operator. (b) An operator A is invertible if and only if Ax = 0 implies x = 0. (c) If an operator A is invertible and vectors x1 , . . . , xn are linearly independent, then Ax1 , . . . , Axn are linearly independent. (d) If operators A and B are invertible, then the operator AB is invertible and we have (AB)−1 = B−1 A−1 . It follows from part (c) in the above theorem that, if E is a finite dimensional vector space and A is a linear invertible operator on E, then R(A) = E. As the following example shows, in infinite dimensional vector spaces it is not necessarily true.

164

Chapter 4

Linear Operators on Hilbert Spaces

Example 4.5.3. Let E = l2 . Define an operator A on E by A(x1 , x2 , . . .) = (0, x1 , x2 , . . .). Clearly, this is a linear invertible operator on l2 whose range is a proper subspace of l2 .
The next example shows that the inverse of a bounded operator is not necessarily bounded.

Example 4.5.4. Let E = l2 . Define an operator A on E by

 xn x2 x3 A(x1 , x2 , . . .) = x1 , , , . . . , , . . . . 2 3 n 

Since

  ∞ ∞ 2       |xn |  2 = (x , x , . . .), A(x1 , x2 , . . .) =  ≤ |x | n 1 2 n2 n=1 n=1

A is a bounded operator. A is also invertible: A−1 (x1 , x2 , . . .) = (x1 , 2x2 , 3x3 , . . . , nxn , . . .). It is easy to see that A−1 is not bounded.




If E is finite dimensional, then the inverse of any invertible operator on E is bounded, because every operator on a finite dimensional space is bounded.

Theorem 4.5.5. Let A be a bounded operator on a Hilbert space H such that

R(A) = H. If A has a bounded inverse, then the adjoint operator A∗ is invertible and (A∗ )−1 = (A−1 )∗ .

Proof: It suffices to show that  −1 ∗ ∗  ∗ A A x = A∗ A−1 x = x

(4.7)

for every x ∈ H. Indeed, for any y ∈ H we have % $ % $  −1 ∗ ∗ % $ −1 A x = A y, A∗ x = AA−1 y, x = y, x y, A and

 %  ∗ % $ ∗ % $ $ y, A∗ A−1 x = Ay, A−1 x = A−1 Ay, x = y, x.

Thus, $  −1 ∗ ∗ % $  ∗  −1 ∗ % A x = y, A A x = y, x y, A which implies (4.7).

for all y ∈ H,


4.5 Normal, Isometric, and Unitary Operators

165

Corollary 4.5.6. If a bounded self-adjoint operator A has a bounded inverse A−1 ,

then A−1 is self-adjoint.

Proof: (A−1 )∗ = (A∗ )−1 = A−1 .




The notion of adjoint operators leads to an important class of operators, called normal operators.

Definition 4.5.7. (Normal operator) A bounded operator T is called a normal operator if it commutes with its adjoint, that is, TT ∗ = T ∗ T . Note that T is normal if and only if T ∗ is normal. Obviously, every selfadjoint operator is normal. The following theorem will help us find examples of normal operators which are not self-adjoint.

Theorem 4.5.8. A bounded operator T is normal if and only if Tx = T ∗ x

for all x ∈ H.

Proof: For all x ∈ H, we have

% $ ∗ T Tx, x = Tx, Tx = Tx 2 .

If T is normal, then we also have 2 % $ % $ %  $ ∗ T Tx, x = TT ∗ x, x = T ∗ x, T ∗ x = T ∗ x , and thus Tx = T ∗ x . Assume now that Tx = T ∗ x for all x ∈ H. By the preceding argument, we have $ ∗ % $ ∗ % TT x, x = T Tx, x for all x ∈ H. Therefore, by Corollary 4.3.8, TT ∗ = T ∗ T.




Note that the condition Tx = T ∗ x for all x ∈ H is much stronger than T = T ∗ .

Example 4.5.9. Let H be a Hilbert space and let Tx = ix for all x ∈ H. Since T ∗ x = −ix = −Tx, T is not self-adjoint. On the other hand, Tx = T ∗ x for all x ∈ H, and thus T is normal.
Theorem 4.5.10. If A is normal, then (α I − A) is normal for any α ∈ C. Proof: Since (α I − A)∗ = (α I − A∗ ), we have (α I − A)(α I − A)∗ = |α|2 I − αA − αA∗ + AA∗ = (α I − A)∗ (α I − A).


166

Chapter 4

Linear Operators on Hilbert Spaces

Theorem 4.5.11. Let T be a bounded operator on a Hilbert space H and let A and B be self-adjoint operators on H such that T = A + iB. Then T is normal if and only if A and B commute. Proof: If T = A + iB, then T ∗ = A − iB, and we have TT ∗ = (A + iB)(A − iB) = A2 + B2 − i(AB − BA)

(4.8)

T ∗ T = (A − iB)(A + iB) = A2 + B2 + i(AB − BA).

(4.9)

and If T is normal, then AB − BA = 0, which proves that A and B commute. On the other hand, if A and B commute, then (4.8) and (4.9) imply ∗ T T = A2 + B2 = TT ∗ .


Theorem 4.5.12. If T is a normal operator, then T n = T n for all n ∈ N. Proof: Note that Theorem 4.2.8 implies that T n ≤ T n for any bounded

operator T. To show that T n ≥ T n we fix an x such that x = 1 and use induction to show that  n  T x ≥ Tx n (4.10) for all n ∈ N. Clearly (4.10) holds for n = 1. If Tx = 0, then the inequality is trivially satisfied for all n ∈ N. Assume now Tx = 0 and that (4.10) holds for n = 1, . . . , m. First note that  2   ∗  $ ∗ % T x = T Tx ≥ T Tx, x = Tx 2 (4.11) by Theorems 4.5.8 and 4.4.14. Now, from (4.11) and the inductive assumption, we obtain      Tx m  m Tx   m+1    T    ≥ Tx T x = Tx T Tx  Tx   m = Tx 1−m T 2 x ≥ Tx 1−m Tx 2m = Tx m+1 .


Definition 4.5.13.

(Isometric operator) A bounded operator T on a Hilbert space H is called an isometric operator if Tx = x for all x ∈ H .

Example 4.5.14. Let (en ), n ∈ N, be a complete orthonormal sequence in a Hilbert space H. There exists a unique operator ∞A such that Aen = en+1 for all n ∈ N. In fact,if x = ∞ α e , then Ax = n n n=1 n=1 αn en+1 . Clearly, A is linear 2 2 and Ax 2 = ∞ |α | = x . Therefore, A is an isometric operator. Opern n=1 ator A is called a one-sided shift operator.


4.5 Normal, Isometric, and Unitary Operators

167

Theorem 4.5.15. A bounded operator T on a Hilbert space H is isometric if and only if T ∗ T = I on H.

Proof: If T is isometric, then for every x ∈ H we have Tx 2 = x 2 , and hence

$

% T ∗ Tx, x = Tx, Tx = Tx 2 = x 2 = x, x

for all x ∈ H. Thus, T ∗ T = I , by Corollary 4.3.8. Similarly, if T ∗ T = I , then

$ % Tx = Tx, Tx = T ∗ Tx, x = x, x = x for all x ∈ H.




Note that isometric operators “preserve inner product,” that is, Tx, Ty = x, y for all x, y ∈ H. In particular, x ⊥ y if and only if Tx ⊥ Ty. An isometric operator is a Hilbert space isomorphism between H and R(T).

Definition 4.5.16.

(Unitary operator) A bounded operator T on a Hilbert space H is called a unitary operator if T ∗ T = TT ∗ = I on H . In the preceding definition it is essential that the domain and the range of T is the entire space H. The following theorem is an immediate consequence of the definition.

Theorem 4.5.17. An operator T is unitary if and only if it is invertible and T −1 = T ∗ .

Example 4.5.18. Let H be the Hilbert space of all sequences of complex num-

bers x = (. . . , x−1 , x0 , x1 , . . .) such that x = product is defined by x, y =

∞ 

∞ 2 −∞ |xn |

< ∞. The inner

xn yn .

−∞

The operator T defined by T(xn ) = (xn−1 ) is a unitary operator. Indeed, T is invertible and Tx, y =

∞  −∞

which implies T ∗ = T −1 .

xn−1 yn =

∞ 

$ % xn yn+1 = x, T −1 y ,

−∞


Note that unitary operators on H can be defined as isometric operators whose range is the entire space H.

168

Chapter 4

Linear Operators on Hilbert Spaces

Example 4.5.19. Let H = L2 ([0, 1]). The operator T on H defined by

(Tx)(t) = x(1 − t) is unitary. It is a one-to-one mapping of H onto H and T(x) = x for all x ∈ H.
A unitary operator is obviously normal, but a normal operator need not be unitary. To see that consider any self-adjoint operator A such that A = 1.

Theorem 4.5.20. Let T be a unitary operator. Then T −1 and T ∗ are unitary. Proof: Note that 

∗ T −1 T −1 = T ∗∗ T −1 = TT −1 = I .

Similarly, T −1 (T −1 )∗ = I , and thus T −1 is unitary. Since T ∗ = T −1 , by Theorem 4.5.17, T ∗ is also unitary.


4.6 Positive Operators First we consider a partial order defined for self-adjoint operators. If A and B are self-adjoint operators on a Hilbert space H, then we write A ≥ B (or B ≤ A) if Ax, x ≥ Bx, x for all x ∈ H. This relation has the following natural properties: If A ≥ B, then −A ≤ −B; If A ≥ B and C ≥ D, then A + C ≥ B + D; If A ≥ 0 and α ≥ 0 (α ∈ R), then αA ≥ 0; If A ≥ B and B ≥ C, then A ≥ C. Proofs of these properties are left as exercises.

Example 4.6.1. Let ϕ and ψ be non-negative continuous function on [a, b] and let A and B be multiplication operators on L2 ([a, b]) defined by Ax = ϕx and Bx = ψx. If ϕ(t) ≥ ψ(t) for every t ∈ [a, b], then A ≥ B. In fact, for any x ∈ L2 ([a, b]), we have 

b

Ax, x = 



a b

≥ a

b

ϕ(t)x(t)x(t)dt =  2 ψ(t)x(t) dt =

 2 ϕ(t)x(t) dt

a



b

ψ(t)x(t)x(t)dt = Bx, x.

a

Theorem 4.6.2. If A is a self-adjoint operator on H and A ≤ 1, then A ≤ I .




4.6 Positive Operators

169

Proof: If A ≤ 1, then Ax, x ≤ A x 2 ≤ x, x = I x, x for all x ∈ H.




Corollary 4.6.3. If A is a self-adjoint operator, then there exists α > 0 (α ∈ R) such that αA ≤ I . Definition 4.6.4. (Positive operator) An operator A is called positive if it is self-adjoint and Ax, x ≥ 0 for all x ∈ H . Clearly, operators A and B in Example 4.6.1 are positive. The next example is a variation of the same idea.

Example 4.6.5. Let K be a positive continuous function defined on [a, b] × [a, b]. The integral operator T on L2 ([a, b]) defined by  (Tx)(s) =

b

K(s, t)x(t)dt a

is positive. Indeed, we have Tx, x =

 b a

b

K(s, t)x(t)x(t)dtds =

a

 b a

b

 2 K(s, t)x(t) dtds ≥ 0

a

for all x ∈ L2 ([a, b]).




Theorem 4.6.6. For any bounded operator A on H, the operators A∗ A and AA∗ are positive.

Proof: For any x ∈ H, we have % $ ∗ A Ax, x = Ax, Ax = Ax 2 ≥ 0 and $

2 % $ %  AA∗ x, x = A∗ x, A∗ x = A∗ x ≥ 0.




Theorem 4.6.7. If A is an invertible positive operator, then its inverse A−1 is positive.

Proof: If y ∈ D(A−1 ), then y = Ax for some x ∈ H, and then % $ −1 % $ −1 A y, y = A Ax, Ax = x, Ax ≥ 0.




170

Chapter 4

Linear Operators on Hilbert Spaces

Example 4.6.8. The product of two positive operators is not necessarily positive. Indeed, consider operators on C2 defined by matrices     1 1 1 0 . and B = A= 1 1 0 0 It is easy to check that both A and B are positive operators, but the product AB is not.


Theorem 4.6.9. Product of two commuting positive operators is a positive operator.

Proof: Let A and B be commuting positive operators. Without loss of generality, we can assume that A = 0. Define a sequence of operators A1 = A/ A

and

An+1 = An − A2n

for n = 1, 2, . . . .

Note that operators An are self-adjoint and commuting. We will show, by induction, that 0 ≤ An ≤ I

(4.12)

for all n ∈ N. For n = 1, (4.12) is satisfied by Theorem 4.6.2. Suppose now (4.12) holds for some k ∈ N. Then % $ % $ % $ 2 Ak (I − Ak )x, x = Ak (I − Ak )x, Ak x = (I − Ak )Ak x, Ak x ≥ 0 and

% $ % $ Ak (I − Ak )2 x, x = Ak (I − Ak )x, (I − Ak )x ≥ 0,

which means A2k (I − Ak ) ≥ 0 and

Ak (I − Ak )2 ≥ 0.

Consequently, Ak+1 = A2k (I − Ak ) + Ak (I − Ak )2 ≥ 0 and

I − Ak+1 = (I − Ak ) + A2k ≥ 0. This shows that (4.12) holds for k + 1, and thus for all n ∈ N, by induction. We have A1 = A21 + A2 = A21 + A22 + A3 = · · · =

n  k=1

A2k + An+1

4.6 Positive Operators

171

and hence n 

A2k = A1 − An+1 ≤ A1 .

k=1

Therefore n 

Ak x, Ak x ≤ A1 x, x.

k=1

 2 This shows that the series ∞ n=1 An x converges and An x → 0. Moreover,  n   2 Ak x = A1 x − An+1 x → A1 x as n → ∞ k=1

or, equivalently, ∞ 

A2n x = A1 x.

n=1

Since B commutes with An for all n ∈ N, we have ABx, x = A BA1 x, x = A

∞  $ 2 % BAn x, x n=1

= A

∞ 

BAn x, An x ≥ 0.




n=1

Corollary 4.6.10. Let A and B be self-adjoint operators. If A ≤ B, then AC ≤ BC for every positive operator C that commutes with both A and B. The next theorem will be useful in studying properties of wavelets (see Chapter 8).

Theorem 4.6.11. Let A be a positive operator on H such that αI ≤ A ≤ β I

(4.13)

for some 0 < α < β. Then (a) A is invertible, (b) R(A) = H, (c)

1 βI

≤ A−1 ≤ α1 I .

Proof: First note that (4.13) is equivalent to α x 2 ≤ Ax, x ≤ β x 2

for all x ∈ H.

(4.14)

172

Chapter 4

Linear Operators on Hilbert Spaces

Thus, Ax = 0 implies x = 0, proving (a). To prove (b) we will prove that the range R(A) is closed and that R(A)⊥ = {0}, and then use Theorem 3.6.9. Consider a sequence yn ∈ R(A) such that yn → y for some y ∈ H. Let yn = Axn for some xn ∈ H. Since $ % α xn − xm 2 ≤ A(xn − xm ), xn − xm = Axn − Axm , xn − xm  = yn − ym , xn − xm  ≤ yn − ym xn − xm , we have α xn − xm ≤ yn − ym , which implies that the sequence (xn ) is a Cauchy sequence, because (yn ) is. Thus (xn ) has a limit in H, say xn → x. From continuity of A, we get yn = Axn → Ax. This implies y = Ax, proving that R(A) is closed. Now assume that Ax, y = 0 for all x ∈ H. But this means that we must have Ay, y = 0, which implies y = 0 by (4.14). Therefore R(A)⊥ = {0}. Finally, if αI ≤ A ≤ β I , then by Theorem 4.6.7 and Corollary 4.6.10, αA−1 ≤ AA−1 ≤ βA−1 and consequently, A−1 ≤

1 I α

and

1 I ≤ A−1 β

proving (c).




The following technical theorem will be used in the proof of Theorem 4.6.14.

Theorem 4.6.12. Let A1 ≤ A2 ≤ · · · ≤ An ≤ · · · be self-adjoint operators on H such that An Am = Am An for all m, n ∈ N. If B is a self-adjoint operator on H such that An B = BAn and An ≤ B for all n ∈ N, then there exists a self-adjoint operator A such that lim An x = Ax

n→∞

for every x ∈ H

4.6 Positive Operators

173

and An ≤ A ≤ B for every n ∈ N.

Proof: Define Cn = B − An . Operators Cn commute with each other and C1 ≥ C2 ≥ · · · ≥ 0. By Theorem 4.6.9, for n > m, the operators (Cm − Cn )Cm

and

Cn (Cm − Cn )

are positive. Hence, % $ % $ 2 Cm x, x ≥ Cm Cn x, x ≥ Cn2 x, x , for every x ∈ H. Since, for an arbitrary fixed x ∈ H, (Ck2 x, x) is a nonincreasing sequence of non-negative numbers, it converges and thus % $ lim Cm Cn x, x = lim Cn2 x, x . m,n→∞

n→∞

Hence, $ % $ 2 % $ % Cm x − Cn x 2 = (Cm − Cn )2 x, x = Cm x, x − 2Cm Cn x, x + Cn2 x, x → 0, as m, n → ∞. Therefore (Cn x) is a Cauchy sequence for every x ∈ H. Consequently, (Cn x), and thus also (An x), are convergent for every x ∈ H. It is easy to check that the operator A defined by Ax = limn→∞ An x is selfadjoint and that An ≤ A ≤ B for every n ∈ N.


Definition 4.6.13.

(Square root) By a square root of a positive operator A we mean a self-adjoint operator B satisfying B2 = A.

Theorem 4.6.14. Every positive operator A has a unique positive square root B. Moreover, B commutes with every operator commuting with A. Proof: Let A ≥ 0 and let α > 0 (α ∈ R) be such that α 2 A ≤ I . Define T0 = 0 and

 1 2 (4.15) α A − Tn2 2 for n = 0, 1, 2, . . . . Note that operators Tn are self-adjoint (as polynomials of A with real coefficients), and they commute with every operator commuting with A. In particular, Tn Tm = Tm Tn for all m and n. Tn+1 = Tn +

174

Chapter 4

Linear Operators on Hilbert Spaces

For every n ∈ N, we have  1 1 I − Tn+1 = (I − Tn )2 + I − α 2 A 2 2

(4.16)

and Tn+1 − Tn =

 1 (I − Tn−1 ) + (I − Tn ) (Tn − Tn−1 ). 2

(4.17)

In view of (4.16), we have Tn ≤ I for all n ∈ N. Moreover, T0 ≤ T1 ≤ · · · ≤ Tn ≤ · · · . Indeed, 1 T1 = α 2 A ≥ 0 = T0 , 2 and if Tn − Tn−1 ≥ 0, then Tn+1 − Tn ≥ 0, by (4.17). By Theorem 4.6.12, the sequence (Tn ) converges to a positive self-adjoint operator T. Letting n → ∞ in (4.15) yields T=T+ that is,



 1 2 α A − T2 , 2

1 T α

2 = A.

Denote B = T/α. The operator B is obviously positive. Since, for each n ∈ N, Tn commutes with every operator commuting with A, so do T and B. It remains to prove the uniqueness. Let C be a positive operator such that C2 = A. Since C commutes with A, C commutes with B. Let x ∈ H and let y0 = (B − C)x. Then $ % By0 , y0  + Cy0 , y0  = (B + C)y0 , y0 $ % = (B + C)(B − C)x, y0 $  % = B2 − C2 x, y0 = 0. Since B and C are positive, we have By0 , y0  = Cy0 , y0  = 0. If D is a positive square root of B, then $ % Dy0 2 = D2 y0 , y0 = By0 , y0  = 0. Hence Dy0 = 0 and also By0 = D(Dy0 ) = 0. In a similar fashion, we can prove that Cy0 = 0. Consequently,

4.7 Projection Operators

175

$ % $ % Bx − Cx 2 = (B − C)2 x, x = (B − C)y0 , x = 0 for arbitrary x ∈ H. This proves B = C.




Definition 4.6.15.

(Strictly positive operator) A self-adjoint operator is called strictly positive or positive definite if Ax, x > 0 for all x ∈ H , x = 0. The operator T in Example 4.6.5 is strictly positive.

4.7 Projection Operators Theorem 3.6.4 states that if S is a closed subspace of a Hilbert space H, then for every x ∈ H, there exists a unique element y ∈ S such that x = y + z and z ∈ S⊥ . Thus every closed subspace induces an operator on H, which assigns to x that unique y.

Definition 4.7.1.

(Orthogonal projection operator) Let S be a closed subspace of a Hilbert space H . The operator P on H defined by Px = y

if x = y + z, y ∈ S, and z ∈ S⊥ ,

(4.18)

is called the orthogonal projection operator onto S, or simply projection onto S. The vector y is called the projection of x onto S. Projection onto a subspace S will be usually denoted by PS . From the uniqueness of the decomposition x = y + z, it follows that projection operators are linear. The Pythagorean formula implies Px 2 = y 2 = x 2 − z 2 ≤ x 2 . Thus, projection operators are bounded and P ≤ 1. The zero operator is a projection operator onto the zero subspace. If PS is a nonzero projection operator, then PS = 1, because for every x ∈ S, we have PS x = x. The identity operator I is the projection operator onto the whole space H.

Example 4.7.2. Let S be a closed subspace of a Hilbert space H, and let {e1 , e2 , . . .} be a complete orthonormal system in S. Then the projection operator PS can be defined by PS x =

∞  x, en en . n=1

176

Chapter 4

Linear Operators on Hilbert Spaces

In particular, if S is of dimension 1 and v ∈ S, v = 1, then PS x = x, vv.




Example 4.7.3. Let H = L2 ([−π, π ]). Every x ∈ H can be represented as x =

y + z, where y is an even function and z is an odd function. The operator defined by Px = y is the projection operator onto the subspace of all even functions. This operator can be also defined as in Example 4.7.2: Px =

∞ 

x, ϕn ϕn ,

n=0

where ϕ0 (t) =

√1 2π

and ϕn (t) =

√1 π

cos nt for n = 1, 2, . . . .




Example 4.7.4. Let H = L2 ([−π, π ]) and let P be an operator on H defined by

 (Px)(t) =

0 if t ≤ 0, x(t) if t > 0.

Then P is the projection operator onto the space of all functions that vanish for t ≤ 0.
Note that if P is a projection on H, then Px, y − Py = 0 for every x, y ∈ H.

(4.19)

Definition 4.7.5.

(Idempotent operator) An operator T is called idempotent if T 2 = T .

Every projection operator is idempotent. Indeed, if P is the projection operator onto a subspace S, then P is the identity operator on S. Since Px ∈ S for every x ∈ H, we have P2 x = P(Px) = Px for all x ∈ H.

Example 4.7.6. Consider the operator T on C2 defined by T(x, y) = (x −y, 0). Obviously, T is idempotent. On the other hand, since $ % T(x, y), (x, y) − T(x, y) = xy − |y|2 , T(x, y) need not be orthogonal to (x, y) − T(x, y), and thus T is not a projection.


Theorem 4.7.7. A bounded operator is a projection if and only if it is idempotent and self-adjoint.

Proof: Let P be a projection operator on H. We have already seen that every projection is idempotent. For any x, y ∈ H, we have Px, y = Px, Py + Px, y − Py = Px, Py

4.7 Projection Operators

177

= Px, Py + x − Px, Py = x, Py, by (4.19). Thus, P is self-adjoint. Assume now that T is an idempotent and self-adjoint operator on H. Define S = {x ∈ H: Tx = x}. Since T is a bounded operator, S is a closed subspace of H. To prove that T is the projection onto S, we need to show that Tx ∈ S and x − Tx ∈ S⊥ for all x ∈ H. The first property follows immediately from the fact that T is idempotent. To prove the second one, note that for any x ∈ H and z ∈ S, we have x − Tx, z = x, z − Tx, z = x, z − x, Tz = x, z − x, z = 0.




Corollary 4.7.8. If P is a projection operator on H, then Px, x = Px 2 for all

x ∈ H.

Proof: From Theorem 4.7.7, we have % $ Px, x = PPx, x = Px, P∗ x = Px, Px = Px 2 .




Example 4.7.9. If PS is the projection operator onto a closed subspace S, then

I − PS is the projection operator onto S⊥ and S⊥ = {x: PS x = 0}. The operator I − PS is sometimes denoted by PS⊥ and called the complementary projection. We thus have PS⊥ = PS⊥ .


In general, the sum of two projection operators is not a projection operator. For example, if P is a nonzero projection operator, then P + P = 2P is not a projection operator, because P + P = 2.

Definition 4.7.10.

(Orthogonality of projection operators) Two projection operators P and Q are called orthogonal if PQ = 0.

Note that for any two projection operators P and Q we have PQ = P∗ Q∗ = (QP)∗ . Thus, PQ = 0 if and only if QP = 0.

Theorem 4.7.11. Two projection operators PR and PS are orthogonal if and only

if R ⊥ S.

Proof: Assume PR PS = 0. If x ∈ R and y ∈ S, then x, y = PR x, PS y = x, PR PS y = 0. Hence R ⊥ S.

178

Chapter 4

Linear Operators on Hilbert Spaces

Now, assume R ⊥ S. If x ∈ H, then PS x ∈ S, and thus PS x ⊥ R. Hence, PR (PS x) = 0 for all x ∈ H. This means that PR and PS are orthogonal.


Theorem 4.7.12. The sum of two projection operators PR and PS is a projection operator if and only if PR PS = 0. In this case, PR + PS = PR⊕S . Proof: If P = PR + PS is a projection operator, then (PR + PS )2 = PR + PS or PR PS + PS PR = 0. Multiplication of this equality by PR from the left yields PR PS + PR PS PR = 0.

(4.20)

Multiplication of the last equality by PR from the right gives 2PR PS PR = 0.

(4.21)

Combining (4.20) and (4.21) yields PR PS = 0. Assume now that PR PS = 0. Then also PS PR = 0, and thus (PR + PS )2 = PR + PS . Therefore, P is idempotent. Since P is also self-adjoint, as a sum of two selfadjoint operators, it is a projection, by Theorem 4.7.7. Let P = PR + PS . For any x ∈ H, we have Px = PR x + PS x ∈ R ⊕ S. Moreover, if x = x1 + x2 with x1 ∈ R, x2 ∈ S, then Px = PR x + PS x = x1 + x2 = x, so P is the identity on R ⊕ S.




Theorem 4.7.13. The product of two projection operators PR and PS is a projection operator if and only if PR and PS commute. In this case, PR PS = PR∩S . Proof: Assume that P = PR PS is a projection. Then P∗ = P and thus PR PS = (PR PS )∗ = PS∗ PR∗ = PS PR . Conversely, if PR PS = PS PR , then (PR PS )∗ = PS∗ PR∗ = PS PR = PR PS ,

4.7 Projection Operators

179

so P = PR PS is self-adjoint. Moreover, P2 = PR PS PR PS = PR2 PS2 = PR PS = P, so P is idempotent. Thus, P is a projection. For x ∈ H, we have Px = PR (PS x) = PS (PR x), and hence Px ∈ R ∩ S. Moreover, for x ∈ R ∩ S, we have Px = PR (PS x) = PR x = x. Consequently, P = PR∩S .


Example 4.7.14. The operators defined by matrices  A=

1 0 0 0



 and

B=

1/2 1/2 1/2 1/2



are projections in C2 . It is easy to check that AB is not a projection.




Theorem 4.7.15. Let R and S be two closed subspaces of a Hilbert space H, and let PR and PS be the respective projections. The following conditions are equivalent: (a) R ⊂ S; (b) PS PR = PR ; (c) PR PS = PR ; (d) PR x ≤ PS x for all x ∈ H.

Proof: Assume R ⊂ S. Then PR x ∈ S for all x ∈ H. Consequently, PS PR x = PR x, and thus (a) implies (b). If PS PR = PR , then PR = PR∗ = (PS PR )∗ = PR∗ PS∗ = PR PS , so (b) implies (c). Assume now PR PS = PR . Then PR x = PR PS x ≤ PR PS x ≤ PS x , for all x ∈ H. Thus (c) implies (d). Finally, suppose (d) holds and (a) does not. Then there exists x ∈ R such / S, we have z = 0 that x ∈ / S. Let x = y + z with y ∈ S and z ∈ S⊥ . Since x ∈ and PR x 2 = y 2 + z 2 > y 2 = PS x 2 , which contradicts (d). Thus, R ⊂ S.




180

Chapter 4

Linear Operators on Hilbert Spaces

4.8 Compact Operators Compact operators constitute an important class of bounded operators. The concept originated from the theory of integral equations of the second kind. They also provide a natural generalization of operators with finitedimensional range.

Definition 4.8.1.

(Compact operator) An operator A on a Hilbert space H is called a compact operator (or completely continuous operator) if, for every bounded sequence (xn ) in H , the sequence (Axn ) contains a convergent subsequence.

Example 4.8.2. Every operator on a finite dimensional Hilbert space is compact. Indeed, if A is an operator on CN , then it is bounded. Therefore, if (xn ) is a bounded sequence, then (Axn ) is a bounded sequence in CN . By the Bolzano–Weierstrass theorem, (Axn ) contains a convergent subsequence.


Example 4.8.3. Let y and z be fixed elements of a Hilbert space H. Define Tx = x, yz. Let (xn ) be a bounded sequence, that is, xn ≤ M for some M > 0 and all n ∈ N. Since   xn , y ≤ xn y ≤ M y , the sequence (xn , y) contains a convergent subsequence (xpn , y). Denote the limit of that subsequence by α. Then Txpn = xpn , yz → αz Therefore, T is a compact operator.

as n → ∞.


Example 4.8.4. Important examples of compact operators are integral operators T on L2 ([a, b]) defined by  b K(s, t)x(t)dt, (Tx)(s) = a

where a and b are finite and K is continuous. We will sketch the proof of compactness of such an operator. Let xn ∈ L2 ([a, b]) and xn ≤ M for n = 1, 2, . . . and some M > 0. Then, by Schwarz’s inequality, we have  b       K(s, t)xn (t)dt ≤ M maxK(s, t) b − a. (Txn )(s) ≤ a

4.8 Compact Operators

181

Thus, the sequence of functions (Txn ) is uniformly bounded. Moreover, for every s1 , s2 ∈ [a, b], we have   (Txn )(s1 ) − (Txn )(s2 ) ≤



b

  K(s1 , t) − K(s2 , t)xn (t)dt

a

+



b

 K(s1 , t) − K(s2 , t)2 dt

≤ a

+



b

 xn (t)2 dt

a



  ≤ M b − a max K(s1 , t) − K(s2 , t). t∈(a,b)

Since K is uniformly continuous, the last inequality implies that the sequence (Txn ) is equicontinuous. Therefore, by Arzela’s theorem (see, for example, Friedman (1982)), (Txn ) contains a uniformly convergent subsequence. This proves that T is compact, because uniform convergence on [a, b] implies convergence in L2 ([a, b]).


Example 4.8.5. Let S be a finite dimensional subspace of a Hilbert space H. The projection operator PS is a compact operator.
Theorem 4.8.6. Compact operators are bounded. Proof: If an operator A is not bounded, then there exists a sequence (xn ) such that xn = 1, for all n ∈ N, and Axn → ∞. Then (Axn ) does not contain a convergent subsequence, which means that A is not compact.
Not every bounded operator is compact.

Example 4.8.7. The identity operator I on an infinite dimensional Hilbert space H is not compact, although it is bounded. In fact, consider an orthonormal sequence (en ) in H. Then the sequence I en = en does not contain a convergent subsequence.
Theorem 4.8.8. The collection of all compact operators on a Hilbert space H is a vector space. The easy proof is left as an exercise.

Theorem 4.8.9. Let A be a compact operator on a Hilbert space H, and let B be a bounded operator on H. Then AB and BA are compact. Proof: Let (xn ) be a bounded sequence in H. Since B is bounded, the sequence (Bxn ) is bounded. Next, since A is compact, the sequence (ABxn ) contains a convergent subsequence, which means that the operator AB is compact. Similarly, since A is compact, the sequence (Axn ) contains a convergent subsequence (Axpn ). Now, since B is bounded (and thus continu-

182

Chapter 4

Linear Operators on Hilbert Spaces

ous), the sequence (BAxpn ) converges. Therefore, the operator BA is compact.
Note that the operators in Examples 4.8.2, 4.8.3, and 4.8.5 have finite dimensional ranges.

Definition 4.8.10.

(Finite-dimensional operator) An operator is called finite-dimensional (or a finite rank operator) if its range is of finite dimension.

Theorem 4.8.11. Finite-dimensional bounded operators are compact. Proof: Let A be a finite dimensional bounded operator, and let {z1 , . . . , zk } be an orthonormal basis of the range of A. Define Tn x = Ax, zn zn for n = 1, . . . , k. Since $ % Tn x = Ax, zn zn = x, A∗ zn zn , the operators Tn are compact as proved in Example 4.8.3. Since A=

k 

Tn ,

n=1

A is compact, by Theorem 4.8.8.




The range of the integral operator considered in Example 4.8.4 is infinite dimensional (unless K = 0). Recall that the space B (H, H) of all bounded operators on a Hilbert space H is a Banach space with respect to the operator norm (see Theorem 1.5.9). It turns out that the space of all compact operators is a closed subspace of B(H, H). As we will see later, this has important consequences.

Theorem 4.8.12. The limit of a uniformly convergent sequence of compact operators is compact. More precisely, if T1 , T2 , . . . are compact operators on a Hilbert space H and Tn −T → 0 as n → ∞ for some operator T on H, then T is compact. Proof: Let (xn ) be a bounded sequence in H. Since T1 is compact, there exists a subsequence (x1,n ) of (xn ) such that (T1 x1,n ) is convergent. Similarly, the sequence (T2 x1,n ) contains a convergent subsequence (T2 x2,n ). In general, for k ≥ 2, let (xk,n ) be a subsequence of (xk−1,n ) such that (Tk xk,n ) is convergent. Consider the sequence (xn,n ). Since it is a subsequence of (xn ),

4.8 Compact Operators

183

we can put xpn = xn,n , where (pn ) is an increasing sequence of positive integers. Obviously, the sequence (Tk xpn ) converges for every k ∈ N. We will show that the sequence (Txpn ) converges too. Let ε > 0. Since Tn −T → 0, there exists k ∈ N such that Tk −T < ε 3M , where M is a constant such that xn ≤ M for all n ∈ N. Next, let k1 ∈ N be such that ε Tk xpn − Tk xpm < 3 for all n, m > k1 . Then Txpn − Txpm ≤ Txpn − Tk xpn + Tk xpn − Tk xpm + Tk xpm − Txpm ε ε ε < + + = ε, 3 3 3 for sufficiently large n and m. Thus, (Txpn ) is a Cauchy sequence in H. Completeness of H implies that (Txpn ) converges.


Corollary 4.8.13. The limit of a convergent sequence of finite-dimensional operators is a compact operator.

Theorem 4.8.14. The adjoint of a compact operator is compact. Proof: Let T be a compact operator on a Hilbert space H and let (xn ) be a bounded sequence in H, that is, xn ≤ M for some M for all n ∈ N. Define yn = T ∗ xn , n = 1, 2, . . . . Since T ∗ is bounded, the sequence (yn ) is bounded. It thus contains a subsequence (ykn ) such that the sequence (Tykn ) converges in H. Now, for any m, n ∈ N, we have 2  ykm − ykn 2 = T ∗ xkm − T ∗ xkn  $ % = T ∗ (xkm − xkn ), T ∗ (xkm − xkn ) $ % = TT ∗ (xkm − xkn ), (xkm − xkn )   ≤ TT ∗ (xkm − xkn ) xkm − xkn ≤ 2M Tykm − Tykn → 0, as m, n → ∞. Therefore (ykn ) is a Cauchy sequence in H, which implies that (ykn ) converges. This proves that T ∗ is a compact operator.
In the next theorem, we characterize compactness of operators in terms of weakly convergent sequences. Recall that we write “xn → x” to denote strong w convergence and “xn → x” to denote weak convergence.

Theorem 4.8.15. An operator T on a Hilbert space H is compact if and only if it maps weakly convergent sequences into strongly convergent sequences. More w precisely, T is compact if and only if xn → x implies Txn → Tx for any xn , x ∈ H.

184

Chapter 4

Linear Operators on Hilbert Spaces w

Proof: Let T be a compact operator. Assume that xn → x and suppose that

Txn → Tx. Then there exists ε > 0 and a subsequence (xpn ) of (xn ) such that Txpn − Tx > ε

(4.22)

for all n ∈ N. Since the sequence (xpn ) is weakly convergent, it is bounded, by Theorem 3.3.14. Compactness of T implies that the sequence (Txpn ) has a strongly convergent subsequence (Txqn ). On the other hand, for every y ∈ H, we have $ % $ % Txn , y = xn , T ∗ y → x, T ∗ y = Tx, y, w

w

so that Txn → Tx, as well as Txqn → Tx. Since we already know that the sequence (Txqn ) is strongly convergent, Theorem 3.3.11 implies Txqn → Tx. But this contradicts (4.22). w Assume now that T is such that Txn → Tx whenever xn → x. Let (zn ) be an arbitrary bounded sequence in H. We want to show that (Tzn ) has a convergent subsequence. Let (en ) be a complete orthonormal sequence in H, and let M be a constant such that zn ≤ M for all n ∈ N. Since   zn , e1  ≤ M, for all n ∈ N, the sequence (zn ) has a subsequence (z1,n ) such that the sequence (z1,n , e1 ) converges. Similarly, since   z1,n , e2  ≤ M, for all n ∈ N, the sequence (z1,n ) has a subsequence (z2,n ) such that the sequence (z2,n , e2 ) converges. Continuing this procedure, we construct sequences (zm,n ), m = 1, 2, 3, . . . , such that (zm+1,n ) is a subsequence of (zm,n ) for every m ∈ N, and the limit limn→∞ zm,n , em  exists for every m ∈ N. Now define xn = zn,n ,

n = 1, 2, . . . .

Clearly, (xn ) is a subsequence of (zn ) and the limit limn→∞ xn , em  exists for every m ∈ N. We will show that (xn ) is weakly convergent. Define αk = lim xn , ek , n→∞

k = 1, 2, . . . .

For any l, n ∈ N we have l ∞       xn , ek 2 ≤ xn , ek 2 = xn 2 ≤ M 2 . k=1

k=1

4.8 Compact Operators

185

By first letting n → ∞, we obtain l 

|αk |2 ≤ M 2 ,

k=1

and then, by letting l → ∞, ∞ 

|αk |2 ≤ M 2 .

k=1

Define z=

∞ 

αk e k .

k=1

Then, for every m ∈ N, ( xn , em  − z, em  = xn , em  −

∞ 

) αk e k , e m

k=1

= xn , em  − αm em , em  = xn , em  − αm → 0,

as n → ∞.

Thus, xn , em  → z, em  as n → ∞ for every m ∈ N. Since span{e1 , e2 , . . .} w is dense in H, it follows that xn → z, by Theorem 3.3.14. Consequently, Txn → Tz.


Corollary 4.8.16. Compact operators map orthonormal sequences into sequences strongly convergent to 0. Proof: Orthonormal sequences are weakly convergent to 0.




Note that, from the preceding theorem it follows that the inverse of a compact operator on an infinite-dimensional Hilbert space, if it exists, is unbounded. It has already been noted that compactness of operators is a stronger condition than boundedness. For operators, boundedness is equivalent to continuity: bounded operators are exactly those operators which map strongly convergent sequences into strongly convergent sequences. Theorem 4.8.15 says that compact operators on a Hilbert space can be characterized as those operators which map weakly convergent sequences into strongly convergent sequences. From this point of view, compactness of operators is a stronger type of continuity. For this reason, compact operators are called sometimes completely continuous operators. The condition has been used by F. Riesz as the definition of compact operators. Hilbert has used still another (equivalent) definition of

186

Chapter 4

Linear Operators on Hilbert Spaces

compact operators: an operator A defined on a Hilbert space H is compact if xn → x weakly and yn → y weakly implies Axn , yn  → Ax, y.

4.9 Eigenvalues and Eigenvectors Concepts discussed in this section play a central role in the theory of operators and their applications. First we need to introduce a number of new ideas.

Definition 4.9.1.

(Eigenvalue) Let A be an operator on a complex vector space E. A complex number λ is called an eigenvalue of A if there is a nonzero vector u ∈ E such that Au = λu.

(4.23)

Every vector u satisfying (4.23) is called an eigenvector of A corresponding to the eigenvalue λ. If E is a function space, eigenvectors are often called eigenfunctions.

Example 4.9.2. Let S be a linear subspace of an inner product space E and let PS be the projection on S. The only eigenvalues of PS are 0 and 1. Indeed, if PS u = λu, for some λ ∈ C and 0 = u ∈ E, then λu = λ2 u, because PS2 = PS . Therefore λ = 0 or λ = 1. The eigenvectors corresponding to 0 are the vectors of E, which are orthogonal to S. The eigenvectors corresponding to 1 are all elements of S.
It is important to note that every eigenvector corresponds to exactly one eigenvalue, but there is always infinitely many eigenvectors corresponding to one eigenvalue. Indeed, every multiple of an eigenvector is an eigenvector. Moreover, several linearly independent vectors may correspond to the same eigenvalue. We have the following simple theorem.

Theorem 4.9.3. The collection of all eigenvectors corresponding to an eigenvalue of an operator is a vector space. The easy proof is left as an exercise.

Definition 4.9.4.

(Eigenspace) The set of all eigenvectors corresponding to an eigenvalue λ is called the eigenspace of λ. The dimension of that space is called the multiplicity of λ. An eigenvalue

4.9 Eigenvalues and Eigenvectors

187

of multiplicity one is called simple or nondegenerate. An eigenvalue of multiplicity greater than one is called multiple or degenerate. In this case the dimension of the eigenspace is also called the degree of degeneracy.

Example 4.9.5. Consider the integral operator A on L2 ([0, 2π ]) defined by 

2π

(Au)(t) =

cos(t − y)u(y)dy.

(4.24)

0

We will show that A has exactly one nonzero eigenvalue λ = π , and its eigenfunctions are u(t) = a cos t + b sin t with arbitrary a and b. The eigenvalue equation is 

2π

(Au)(t) =

cos(t − y)u(y)dy = λu(t)

0

or



2π

cos t



2π

cos yu(y)dy + sin t

0

sin yu(y)dy = λu(t).

(4.25)

0

This means that, for λ = 0, u is a linear combination of cosine and sine functions, that is, u(t) = a cos t + b sin t,

(4.26)

where a, b ∈ C. Substituting this into (4.25), we obtain π a = λa

and

π b = λb.

(4.27)

Hence, λ = π , which means that A has exactly one nonzero eigenvalue and its eigenfunctions are given by (4.26). This is a two-dimensional eigenspace, so the multiplicity of the eigenvalue is 2. Equation (4.25) reveals that λ = 0 is also an eigenvalue of A. The corresponding eigenfunctions are all functions orthogonal to cos t and sin t. Therefore, λ = 0 is an eigenvalue of infinite multiplicity.
Note that, if λ is not an eigenvalue of A, then the operator A − λI is invertible, and conversely. If space E is finite dimensional and λ is not an eigenvalue of A, then the operator (A − λI )−1 is bounded, because all operators on a finite dimensional space are bounded. The situation for infinite dimensional spaces is more complicated.

188

Chapter 4

Linear Operators on Hilbert Spaces

Definition 4.9.6.

(Resolvent and spectrum) Let A be an operator on a normed space E. The operator Aλ = (A − λI )−1

is called the resolvent of A. The values of λ for which Aλ is defined on the whole space E and is bounded are called regular values of A. The set of all regular values of A is called the resolvent set and is denoted by ρ(A). The complement of ρ(A) in C is called the spectrum of A and is denoted by σ (A). The spectral radius of A, denoted by r(A), is defined by
 r(A) = sup |λ|: λ ∈ σ (A) .

Every eigenvalue belongs to the spectrum. The following example shows that the spectrum may contain points that are not eigenvalues. In fact, a nonempty spectrum may contain no eigenvalues at all.

Example 4.9.7. Let E be the space C ([a, b]) of continuous functions on the interval [a, b]. For a fixed u ∈ C ([a, b]), consider the operator A defined by (Ax)(t) = u(t)x(t). Since (A − λI )−1 x(t) =

x(t) , u(t) − λ

the spectrum of A consists of all λ’s such that λ − u(t) = 0 for some t ∈ [a, b]. This means that the spectrum of A is exactly the range of u. If u(t) = c is a constant function, then λ = c is an eigenvalue of A. On the other hand, if u is a strictly increasing function, then A has no eigenvalues. The spectrum of A in such a case is the interval σ (A) = [u(a), u(b)].
The problem of finding eigenvalues and eigenvectors is called the eigenvalue problem. One of the main sources of eigenvalue problems in mechanics is the theory of oscillating systems. The state of a given system at a given time t may be represented by an element u(t) ∈ H, where H is an appropriate Hilbert space of functions. The equation of motion in classical mechanics is d2 u = Au, dt 2

(4.28)

where A is an operator in H. If the system oscillates, the time dependence of u is sinusoidal, so that u(t) = v sin ωt, where v is a fixed element of H. If A is

4.9 Eigenvalues and Eigenvectors

189

linear, then (4.28) becomes   Av = −ω2 v.

(4.29)

This means that −ω2 is an eigenvalue of A.

Theorem 4.9.8. If A is a bounded linear operator in a Banach space E and A < |λ|, then Aλ = (A − λI )−1 is a bounded operator, ∞  An , λn+1 n=0

(4.30)

1 . |λ| − A

(4.31)

Aλ = − and Aλ ≤

Proof: Since A/λ < 1, we have ∞  n ∞  n  A   A  ≤   < ∞.  λn  λ n=0

n=0

Therefore, by the completeness of B (E, E), there exists a bounded linear operator B on E such that ∞  An

B=

n=0

λn

.

Moreover, (A − λI )B = (A − λI )

 ∞  An n=0

=

λn

∞  An+1 − λAn n=0

λn

=

=λ

∞  An (A − λI ) n λ n=0

∞  n+1  A n=0

λn+1

−

An λn

 = −λI .

Similarly, B(A − λI ) = −λI . Thus, Aλ = (A − λI )

−1

∞  B An =− =− . λ λn+1 n=0

To prove (4.31), we observe that ∞  n  1 1  1 A = 1 Aλ ≤ = .   |λ| n=0 λ |λ| 1 − A/λ |λ| − A




190

Chapter 4

Linear Operators on Hilbert Spaces

The representation (4.30) is usually referred to as the Neumann series (Carl Gottfried (1832–1925)). Note that, if A < 1, then we have (I − Neumann n A)−1 = ∞ A . n=0

Corollary 4.9.9. If A is a bounded operator on a Banach space, then r(A) ≤ A . If the eigenvalues are considered as points in the complex plane, the above result implies that all eigenvalues of a bounded operator A lie in the closed disk of radius A centered at the origin.

Theorem 4.9.10. Let T be an invertible operator on a vector space E, and let A

be an operator on E. The operators A and TAT −1 have the same eigenvalues.

Proof: Let λ be an eigenvalue of A. This means that there exists a nonzero vector u such that Au = λu. Since T is invertible, Tu = 0 and TAT −1 (Tu) = TAu = T(λu) = λTu. Thus, λ is an eigenvalue of TAT −1 . Assume now that λ is an eigenvalue of TAT −1 , that is, TAT −1 u = λu for some nonzero vector u = Tv. Since AT −1 u = λT −1 u and T −1 u = 0, λ is an eigenvalue of A.
The remaining theorems of this section describe properties of operators on Hilbert spaces.

Theorem 4.9.11. All eigenvalues of a self-adjoint operator on a Hilbert space are real.

Proof: Let λ be an eigenvalue of a self-adjoint operator A, and let u be an eigenvector of λ, u = 0. Then λu, u = λu, u = Au, u = u, Au = u, λu = λu, u. Since u, u = 0, we conclude λ = λ.




Theorem 4.9.12. If A is a bounded self-adjoint operator on a Hilbert space, then r(A) = A .

Proof: The theorem is obviously true for A = 0. Let A be a nonzero, bounded, self-adjoint operator on a Hilbert space H. In view of Corollary 4.9.9, it suffices to show that there exists a λ ∈ σ (A) such that |λ| = A . Since, by Theorem 4.4.14, A = sup x =1 |Ax, x|, there exists a sequence xn ∈ H such that xn = 1 and |Axn , xn | → A , as n → ∞. Without loss of generality, we may assume that Axn , xn  → λ, where |λ| = A . For every n ∈ N, we have Axn − λxn 2 = Axn 2 − 2λAxn , xn  + λ2 xn 2   ≤ A 2 − 2λAxn , xn  + λ2 = 2λ λ − Axn , xn  .

191

4.9 Eigenvalues and Eigenvectors

Thus, Axn − λxn → 0,

(4.32)

as n → ∞. Now, suppose λ ∈ ρ(A). Then, from (4.32) and continuity of (A − λI )−1 , we obtain   1 = xn = (A − λI )−1 (A − λI )xn  → 0, as n → ∞. This contradiction shows that λ ∈ σ (A).




Theorem 4.9.13. All eigenvalues of a positive operator are non-negative. All eigenvalues of a strictly positive operator are positive.

Proof: Let A be a positive operator, and let Ax = λx for some x = 0. Since A is self-adjoint, we have 0 ≤ Ax, x = λx, x = λ x 2 .

(4.33)

Thus, λ ≥ 0. The proof of the second part of the theorem is obtained by replacing “≤” by “ n0 . For every x ∈ H and every k, m ∈ N such that n0 < k < m, we have 2  m m m       λn Pn x = λn Pn x 2 = |λn |2 Pn x 2 (4.35)    n=k

n=k

≤ε

2

m  n=k

n=k

2  m  m     2   2 Pn x = ε  Pn x ≤ ε  Pn  x 2 ,     2

2

n=k

n=k

(4.36) where the first and the last equalities follow from orthogonality of projections Pn . The sum m n=k Pn , being a finite sum of projection operators, is a projection operator and its operator norm is 1. Thus, (4.35) yields  m    m         = sup λ P λ P x   n n n n  ≤ ε,    x =1 n=k

n=k

 for any n0 < k < m, proving that the sequence of partial sums m n=k λn Pn is a Cauchy sequence. (b) Denote the range of Pn by R(Pn ) and let n0 ∈ N. If u ∈ R(Pn0 ), then Pn0 u = u and Pn u = 0 for all n = n0 , because the projections Pn are orthogonal. Thus, ∞  Au = λn Pn u = λn0 u, n=1

which shows that λn0 is an eigenvalue of A. To prove that there are no other nonzero eigenvalues, suppose u is an eigenvector corresponding to an eigenvalue λ. Set vn = Pn u, n = 1, 2, . . . , and let w = Qu, where Q is the projection on the orthogonal complement

4.9 Eigenvalues and Eigenvectors

195

of R(A). Then u=

∞ 

vn + w

(4.37)

n=1

with w ⊥ R(Pn ) for all n ∈ N. Clearly, ∞  ∞  ∞    A vn + w = A vn = Avn n=1

n=1

n=1

since Pn w = 0 and A is continuous. Consequently, the eigenvalue equation can be written in the form ∞  ∞   λn vn = λ vn + w n=1

n=1

or ∞  (λ − λn )vn + λw = 0.

(4.38)

n=1

Since all vectors in (4.38) are orthogonal, the sum vanishes only if every term vanishes. Hence, λw = 0, and for every n ∈ N either λ = λn or vn = 0. Finally, if u in (4.37) is a nonzero eigenvector, then either w = 0 or vk = 0 for some k ∈ N. Therefore, λ = 0 or λ = λk for some k ∈ N, by (4.38). (c) Suppose all λn ’s are real. Since projections are self-adjoint operators, for any x, y ∈ H, we have ∞ ∞   Ax, y = λn Pn x, y = λn Pn x, y n=1

=

∞  n=1

n=1

λn x, Pn y =

∞  x, λn Pn y = x, Ay. n=1

(d) If all projections Pn are finite-dimensional, then A is compact by Corollary 4.8.13.
In some applications, the following generalization of the notion of eigenvalue is useful.

Definition 4.9.22.

(Approximate eigenvalue) Let T be an operator on a Hilbert space H . A scalar λ is called an approximate eigenvalue of T if there exists a sequence of vectors xn ∈ H such that xn = 1 for all n ∈ N and Txn − λxn → 0 as n → ∞.

196

Chapter 4

Linear Operators on Hilbert Spaces

Obviously, every eigenvalue is an approximate eigenvalue.

Example 4.9.23. Let (en ) be a orthonormal sequence in a Hilbert space H. Let

(λn ) be a sequence of scalars such that λn → λ and λn = λ for all n ∈ N. Define an operator T on H by Tx =

∞ 

λn x, en en .

n=1

It is easy to see that every λn is an eigenvalue of T, but λ is not. On the other hand,   Ten − λen = λn en − λen = (λn − λ)en  = |λn − λ| → 0 as n → ∞. Thus, λ is an approximate eigenvalue of T.




Theorem 4.9.24. If T is a compact operator, then every nonzero approximate eigenvalue of T is an eigenvalue. Proof: Let (xn ) be a sequence of vectors such that xn = 1 for all n ∈ N and Txn − λxn → 0 as n → ∞, for some λ = 0. Since T is compact, there exists a subsequence (xpn ) of (xn ) such that Txpn → y as n → ∞, for some y. Then y − λxpn ≤ y − Txpn + Txpn − λxpn → 0, as n → ∞. Since λ = 0, we have xpn → λy . If we let u = λy , then u = 1 and Tu − λu ≤ Tu − Txpn + Txpn − y → 0, as n → ∞. Thus, Tu = λu.




For further properties of approximate eigenvalues see the exercises at the end of this chapter.

4.10 Spectral Decomposition Let H be a finite-dimensional Hilbert space, say H = CN . It is known from linear algebra that eigenvectors of a self-adjoint operator on H form an orthogonal basis of H. The following theorems generalize this result to infinitedimensional spaces.

Theorem 4.10.1. (Hilbert–Schmidt theorem) For every compact, self-adjoint operator A on an infinite-dimensional Hilbert space H, there exists an orthonormal system of eigenvectors (un ) corresponding to nonzero eigenvalues (λn ) such

197

4.10 Spectral Decomposition

that every element x ∈ H has a unique representation in the form x=

∞ 

αn un + v,

(4.39)

n=1

where αn ∈ C and v ∈ N (A).

Proof: By Theorem 4.9.16 and Corollary 4.9.17, there exists an eigenvalue λ1 of A such that

  |λ1 | = sup Ax, x. x ≤1

Let u1 be a normalized eigenvector corresponding to λ1 . We set Q1 = {x ∈ H: x ⊥ u1 }, that is, Q1 is the orthogonal complement of the set {u1 }. Thus, Q1 is a closed linear subspace of H. If x ∈ Q1 , then Ax, u1  = x, Au1  = λ1 x, u1  = 0, which means that x ∈ Q1 implies Ax ∈ Q1 . Therefore, A maps the Hilbert space Q1 into itself. We can again apply Theorem 4.9.16 and Corollary 4.9.17 with Q1 in place of H. This gives an eigenvalue λ2 such that 
  |λ2 | = sup Ax, x: x ∈ Q1 . x ≤1

Let u2 be a normalized eigenvector of λ2 . Clearly u1 ⊥ u2 . Next we set Q2 = {x ∈ Q1 : x ⊥ u2 }, and repeat the above argument. Having eigenvalues λ1 , . . . , λn and the corresponding normalized eigenvectors u1 , . . . , un , we define Qn = {x ∈ Qn−1 : x ⊥ un } and choose an eigenvalue λn+1 such that 
  |λn+1 | = sup Ax, x: x ∈ Qn . x ≤1

(4.40)

For un+1 we choose a normalized vector corresponding to λn+1 . This procedure can terminate after a finite number of steps. Indeed, it can happen that there is a positive integer k such that Ax, x = 0 for every x ∈ Qk . Then every element x of H has a unique representation x = α1 u1 + · · · + αk uk + v,

198

Chapter 4

Linear Operators on Hilbert Spaces

where Av = 0, and Ax = λ1 α1 u1 + · · · + λk αk uk , which proves the theorem in this case. Now suppose that the described procedure yields an infinite sequence of eigenvalues (λn ) and eigenvectors (un ). Let S be the closed space spanned by the vectors u1 , u2 , . . . , that is, ∞  ∞   2 S= αn un : |αn | < ∞ . n=1

n=1

By Theorem 3.6.6, every x ∈ H has a unique decomposition x = u + v or x=

∞ 

αn un + v,

n=1

where v ∈ S⊥ . It remains to prove that Av = 0 for all v ∈ S⊥ . Let v ∈ S⊥ , v = 0. Define w = v/ v . Then Av, v = v 2 Aw, w. Since w ∈ S⊥ ⊂ Qn for every n ∈ N, by (4.40), we have     
  Av, v = v 2 Aw, w ≤ v 2 sup Ax, x: x ∈ Qn x ≤1

= v 2 |λn+1 | → 0. This implies Av, v = 0 for every v ∈ S⊥ . Therefore, by Theorem 4.4.14, the norm of A restricted to S⊥ is 0, and thus Av = 0 for all v ∈ S⊥ .


Corollary 4.10.2. (Spectral theorem for compact self-adjoint operators) Let A be a compact, self-adjoint operator on an infinite-dimensional Hilbert space H. Then H has a complete orthonormal system (an orthonormal basis) {v1 , v2 , . . .} consisting of eigenvectors of A. Moreover, for every x ∈ H, Ax =

∞ 

λn x, vn vn ,

(4.41)

n=1

where λn is the eigenvalue corresponding to vn .

Proof: To obtain a complete orthonormal system {v1 , v2 , . . .}, we need to complement the system {u1 , u2 , . . .}, defined in the proof of Theorem 4.10.1, with an arbitrary orthonormal basis of N (A). The eigenvalues corresponding to the vectors that form N (A) are all equal zero. Equality (4.41) follows from the continuity of A.


4.10 Spectral Decomposition

199

Corollary 4.10.3. Let A be a compact, self-adjoint operator on a Hilbert space H. Then A=

∞ 

λn Pn ,

(4.42)

n=1

where Pn is a projection on finite-dimensional subspace of H.

Proof: Let {v1 , v2 , . . .} be a complete orthonormal system of eigenvectors of

A corresponding to eigenvalues {λ1 , λ2 , . . .}. Let Pn be the projection operator onto the one-dimensional space spanned ∞ by vn . Since Pn x = x, vn vn , equation (4.41) can be written as Ax = n=1 λn Pn x. Here is another way of representing A in the form (4.42). Let {λ1 , λ2 , . . .} be all distinct nonzero eigenvalues of A and let Pn be the projection onto the eigenspace corresponding to λn . By Theorem 4.9.18, those eigenspaces are finite-dimensional.
Corollary 4.10.3 is just another version of the spectral theorem. This version is important because it has natural extensions to more general classes of operators. It is also useful because it leads to an elegant expression for powers and other functions of an operator. Let A, λn , and Pn be as in Corollary 4.10.3. Then A =A 2

∞ 

 λn Pn =

n=1

∞ 

λn APn =

n=1

∞ 

λ2n Pn ,

n=1

because APn x = λn Pn x for every x ∈ H. Similarly, for any k ∈ N, we get A = k

∞ 

λkn Pn ,

(4.43)

n=1

and, hence, for any polynomial p(t) = αn t n + · · · + α1 t, we have p(A) =

∞ 

p(λn )Pn .

n=1

The constant term in p must be zero, because otherwise the sequence (p(λn )) would not converge to zero. To deal with polynomials with a nonzero constant term α0 , we add α0 I to the series. Note that, in such a case, p(A) is not a compact operator. This method can be generalized in the following way.

200

Chapter 4

Linear Operators on Hilbert Spaces

Definition 4.10.4.

(Function of an operator) Let f be a real-valued function on R such that f (λ) → 0

as λ → 0 and f (0) = 0.  For a compact, self-adjoint operator A = ∞ n=1 λn Pn , we define f (A) =

∞ 

f (λn )Pn .

(4.44)

n=1

Theorem 4.9.21 ensures that the series in (4.44) converges and that f (A) is self-adjoint and compact.  Example 4.10.5. Let A = ∞ n=1 λn Pn be a compact, self-adjoint operator such that λn ≥ 0 for all n ∈ N. For any α > 0, we can define Aα by Aα x =

∞ 

λαn Pn x.

n=1

Note that in the case α = 12 , the defined operator A1/2 is the same as the square root of A introduced in Definition 4.6.13. Indeed, by (4.43), we have ∞ ∞   2 √ 2  A = λn Pn = λn Pn = A, n=1

n=1

because all λn ’s are non-negative.
∞ Example 4.10.6. Let A = n=1 λn Pn be a compact, self-adjoint operator. We can define sine of A by sin A =

∞ 

(sin λn )Pn .

n=1

The condition that f (λ) → 0 as λ → 0 in Definition 4.10.1 can be replaced ∞ by boundedness of f in a neighborhood of the origin. Indeed, if A = n=1 λn Pn and Pn x = x, vn vn , then, for any x ∈ H, we have ∞    f (λn )x, vn vn , f (A) x = n=1

where convergence of the series is justified by Theorem 3.4.10, because     f (λn )x, vn 2 ≤ M x, vn 2

4.11 Unbounded Operators

201

for some constant M, and hence (f (λn )x, λn ) ∈ l2 . Clearly, in this case, we cannot expect f (A) to be a compact operator.


Theorem 4.10.7. If eigenvectors u1 , u2 , . . . of a self-adjoint operator T on a Hilbert space H form a complete orthonormal system in H and all eigenvalues are positive (or non-negative), then T is strictly positive (or positive). Proof: Suppose u1 , u2 , . . . is a complete orthonormal system of eigenvalues

of T corresponding to eigenvalues  λ1 , λ2 , . . . . Then any nonzero vector u ∈ H can be represented as u = ∞ n=1 αn un , and we have ) ( ∞ ∞ ∞    αn un = αn Tu, un  = αn u, Tun  Tu, u = Tu, n=1

=

∞ 

αn u, λn un  =

n=1

=

∞ 

n=1 ∞ 

n=1

λn αn u, un  =

n=1

∞ 

λn α n α n

n=1

λn |αn |2 ≥ 0

n=1

if all eigenvalues are non-negative. If all λn ’s are positive, then the last inequality becomes strict.


Theorem 4.10.8. For any two commuting, compact, self-adjoint operators A and B on a Hilbert space H, there exists a complete orthonormal system in H of common eigenvectors of A and B. Proof: Let λ be an eigenvalue of A, and let S be the corresponding eigenspace. For any x ∈ S, we have ABx = BAx = B(λx) = λBx. This means that Bx is an eigenvector of A corresponding to λ provided Bx = 0. In any case, Bx ∈ S and hence B maps S into itself. Since B is a compact, self-adjoint operator, by Corollary 4.10.2, S has an orthonormal basis consisting of eigenvalues of B, but these vectors are also eigenvectors of A, because they belong to S. If we repeat the same with every eigenspace of A, then the union of all these eigenvectors will be an orthonormal basis of H.


4.11 Unbounded Operators Boundedness of an operator was an essential assumption in almost every theorem proved in this chapter. Methods used were developed with boundedness

202

Chapter 4

Linear Operators on Hilbert Spaces

or continuity in mind. However, in the most important applications of the theory of Hilbert spaces we often have to deal with operators which are not bounded. In this section, we will briefly discuss some basic problems, concepts, and methods in the theory of unbounded operators. Some concepts, like the adjoint operator, will require a new approach in this setting. An operator A defined in a Hilbert space H, that is, R(A) ⊂ H, is called unbounded if it is not bounded. Therefore, to show that an operator A is unbounded it suffices to find a sequence of elements xn ∈ H such that xn ≤ M (for some M and all n ∈ N) and Axn → ∞. Since for linear operators boundedness is equivalent to continuity, unboundedness is equivalent to lack of continuity (at any point). Consequently, we can show that an operator A is unbounded by finding a sequence (xn ) convergent to 0 such that the sequence (Axn ) does not converge to 0. One of the most important unbounded operators is the differential operator on L2 (R) (see Example 4.2.3). Other important unbounded operators arise from quantum mechanics and will be discussed in Chapter 7.

Example 4.11.1. Let A be a compact operator on an infinite dimensional Hilbert space H. If A is invertible, then A−1 is unbounded. Indeed, let vn ∈ H be an orthonormal sequence and let zn = Avn . Then zn → 0, but A−1 zn → 0.
It will be convenient to adopt the following convention: When we say “A is an operator on a Hilbert space H,” we mean that the domain of A is the whole space H, and when we say “A is an operator in a Hilbert space H,” we mean that the domain of A is a subset of H. If the domain of a bounded operator A is a proper subspace of a Hilbert space H, then A has a unique extension to a bounded operator defined on the closure of D (A). More precisely, there exists a bounded operator B defined on cl D (A) such that Ax = Bx for every x ∈ D (A). Such an operator can be defined by Bx = lim Axn , n→∞

where xn ∈ D (A) and xn → x.

In this case B = A . Then B can be extended to a bounded operator C defined on all of H by C = BPD(B) , where PD(B) is the orthogonal projection operator onto the subspace D (B). Again we have C = B . We may thus always assume that the domain of a bounded operator is the entire space H, or at least a closed subspace of H. An unbounded operator defined on a proper subspace of a Hilbert space usually does not have a natural extension onto the closure of its domain. For instance, the differential operator is defined on a dense subspace of L2 (R), but it does not have a natural extension onto L2 (R). On the other hand, it may be still

4.11 Unbounded Operators

203

possible to extend the domain of an unbounded operator in such a way that, although the domain of the extension is not the whole space, it has better properties.

Definition 4.11.2.

(Extension of operators) Let A and B be operators in a vector space E. If D(A) ⊂ D(B)

and Ax = Bx

for every x ∈ D(A),

then B is called an extension of A and we write A ⊂ B. When performing typical operations on operators defined on subspaces of a common vector space E, we have to keep track of the domains. For instance, the operator A + B is defined for all x ∈ D (A) ∩ D (B), that is, D (A + B) = D(A) ∩ D(B). It may happen that D(A) ∩ D(B) = {0}, and then the sum A + B does not make much sense. Similarly, D (AB) = {x ∈ D (B): Bx ∈ D (A)}. The usual properties need not hold. For example, we have the equality (A + B)C = AC + BC, but, in general, the inclusion AB + AC ⊂ A(B + C) cannot be replaced by equality.

Definition 4.11.3.

(Densely defined operator) An operator A defined in a normed space E is called densely defined if its domain is a dense subset of E, that is, cl D(A) = E.

d The differential operator D = dx is densely defined in L2 (R), because the subspace of differentiable functions is dense in L2 (R).

Theorem 4.11.4. Let A be a densely defined operator in a Hilbert space H and let E be the set of all y ∈ H for which Ax, y is a continuous functional on D (A). There exists a unique operator B defined on E such that Ax, y = x, By

for all x ∈ D (A) and y ∈ E.

Proof: For any y ∈ E, the functional fy (x) = Ax, y, being continuous on a dense subspace of H, has a unique extension to a continuous functional f˜y on H. By the Riesz representation theorem, there exists a unique zy ∈ H such that f˜y (x) = x, zy  for all x ∈ H. If we define B(y) = zy , then we will have Ax, y = fy (x) = f˜y (x) = x, zy  = x, By

204

Chapter 4

Linear Operators on Hilbert Spaces

for all x ∈ D (A) and y ∈ E. Proof of linearity of B is left as an exercise.




The preceding result justifies the following definition.

Definition 4.11.5.

(Adjoint of a densely defined operator) Let A be a densely defined operator in a Hilbert space H . The adjoint A∗ of A is the operator defined on the set of all y ∈ H for which Ax, y is a continuous functional on D(A) and such that $ %   Ax, y = x, A∗ y for all x ∈ D (A) and y ∈ D A∗ .

Note that for a bounded operator on H this definition is equivalent to Definition 4.4.1.

Example 4.11.6. Let C01 (R) denote the space of all continuously differentiable

functions on R with compact support. This is a dense subspace of L2 (R). Consider the differential operator D defined on C01 (R). Since     ∞  ∞ d d x(t) y(t)dt = − Dx, y = x(t) y(t) dt, dt −∞ dt −∞ Dx, y is a continuous functional on C01 (R) (with respect to the norm of L2 (R)) for every y ∈ L2 (R) such that y ∈ L2 (R). Moreover,     ∞ d d x(t) y(t) dt = x(t) − y(t) dt. Dx, y = − dt dt −∞ −∞ 

∞



Note that it is not correct to write D∗ = −D, since the domain of D∗ is not C01 (R).


Theorem 4.11.7. Let A and B be densely defined operators in a Hilbert space H. (a) If A ⊂ B, then B∗ ⊂ A∗ . (b) If D (B∗ ) is dense in H, then B ⊂ B∗∗ .

Proof: To prove (a), consider a y ∈ D(B∗ ). Then Bx, y, as a function of

x, is a continuous functional on D (B). Since D (A) ⊂ D (B), Bx, y is a continuous functional on D (A). But Bx = Ax for x ∈ D (A), so Ax, y is a continuous functional on D (A). This proves that y ∈ D (A∗ ). The equality A∗ y = B∗ y for y ∈ D (B∗ ) follows from uniqueness of the adjoint operator. To prove (b), observe that the condition %   $ Bx, y = x, B∗ y for all x ∈ D (B) and all y ∈ D B∗

4.11 Unbounded Operators

can be rewritten as $ ∗ % B y, x = y, Bx

  for all y ∈ D B∗ and all x ∈ D (B).

Since D (B∗ ) is dense in H, B∗∗ exists and we have     $ ∗ % $ ∗∗ % B y, x = y, B x for all y ∈ D B∗ and all x ∈ D B∗∗ .

205

(4.45)

(4.46)

Now, by an argument similar to that in part (a), we can show that D (B) ⊂ D(B∗∗ ) and B(x) = B∗∗ (x) for any x ∈ D(B).


Theorem 4.11.8. If A is a one-to-one operator in a Hilbert space and both A and its inverse A−1 are densely defined, then A∗ is also one-to-one and  ∗ −1  −1 ∗ = A . A

(4.47)

Proof: Let y ∈ D(A∗ ). Then for every x ∈ D(A−1 ), we have A−1 x ∈ D(A) and hence

% $ % $ −1 A x, A∗ y = AA−1 x, y = x, y.

This means that A∗ y ∈ D ((A−1 )∗ ) and  ∗  −1 ∗ ∗ A y = AA−1 y = y. A

(4.48)

Next, take an arbitrary y ∈ D (A−1 )∗ . Then, for each x ∈ D (A), we have Ax ∈ D (A−1 ). Hence % ∗ % $ $  Ax, A−1 y = A−1 Ax, y = x, y. This shows that (A−1 )∗ y ∈ D (A∗ ) and  ∗  ∗ A∗ A−1 y = A−1 A y = y.

(4.49)

Now equality (4.47) follows from (4.48) and (4.49).




Theorem 4.11.9. If A, B, and AB are densely defined operators in H, then B∗ A∗ ⊂ (AB)∗ .

Proof: Suppose x ∈ D(AB) and y ∈ D(B∗ A∗ ). Since x ∈ D(B) and A∗ y ∈ D(B∗ ), it follows that

% $ % $ Bx, A∗ y = x, B∗ A∗ y .

On the other hand, since Bx ∈ D (A) and y ∈ D (A∗ ), we have % $ ABx, y = Bx, A∗ y . Hence

% $ ABx, y = x, B∗ A∗ y .

206

Chapter 4

Linear Operators on Hilbert Spaces

Since this holds for all x ∈ D (AB), we have y ∈ D ((AB)∗ ) and (B∗ A∗ )y = (AB)∗ y.
Self-adjoint operators have been already discussed in Section 4.4. In that section we limited our discussion to bounded operators. Without the boundedness condition the matter is more delicate.

Definition 4.11.10.

(Self-adjoint operator) Let A be a densely defined operator in a Hilbert space H . A is called self-adjoint if A = A∗ . Remember that A = A∗ means that D (A∗ ) = D (A) and A(x) = A∗ (x) for all x ∈ D (A). If A is a bounded, densely defined operator in H, then A has a unique extension to a bounded operator on H, and then its domain as well as the domain of its adjoint, is the whole space H. In the case of unbounded operators, it is possible that a densely defined operator A has an adjoint A∗ such that A(x) = A∗ (x) whenever x ∈ D (A) ∩ D (A∗ ), but D (A∗ ) = D (A), and thus A is not self-adjoint. The following relaxation of the conditions in Definition 4.11.10 seems reasonable in the case of unbounded operators.

Definition 4.11.11.

(Symmetric operator) A densely defined operator A in a Hilbert space H is said to be symmetric if Ax, y = x, Ay

for every x, y ∈ D(A).

Clearly, every self-adjoint operator is symmetric. In the first of the following two examples, we present an unbounded self-adjoint operator. The second one shows that a symmetric operator need not be self-adjoint.

Example 4.11.12. Consider a bounded operator on H = l2 defined by 

 xn A(xn ) = . n

Note that A is self-adjoint and one-to-one. The subspace R(A) = D (A−1 ) consists of all sequences (yn ) ∈ l2 such that ∞ 

n2 |yn |2 < ∞

n=1

and it is dense in H. The inverse A−1 is defined by A−1 (yn ) = (nyn ).

207

4.11 Unbounded Operators

Clearly, A−1 is an unbounded operator. By Theorem 4.11.8,  −1 ∗  ∗ −1 = A = A−1 . A Thus, A−1 is self-adjoint.




Example 4.11.13. Consider the operator A = id/dt with the domain
   D(A) = f ∈ L2 [a, b] : f is continuous and f (a) = f (b) = 0 . Since



b

Af , g =

if (t)g(t)dt =

a



b

f (t)ig (t)dt = f , Ag

a

for all f , g ∈ D (A), A is symmetric. However, Af , g is a continuous functional on D (A) for any function g continuously differentiable on [a, b], not necessarily satisfying g(a) = g(b). Consequently, D (A∗ ) = D (A) and A is not self-adjoint.
The following theorem gives a nice characterization of symmetric operators.

Theorem 4.11.14. A densely defined operator A in a Hilbert space H is symmetric if and only if A ⊂ A∗ . Proof: Suppose A ⊂ A∗ . Since $ % Ax, y = x, A∗ y

  for all x ∈ D (A) and all y ∈ D A∗ ,

(4.50)

we have Ax, y = x, Ay for all x, y ∈ D (A).

(4.51)

Thus, A is symmetric. If A is symmetric, then (4.50) and (4.51) hold, which implies A ⊂ A∗ .




Corollary 4.11.15. If A is symmetric, then A∗ is the maximal symmetric extension of A.

Proof: Let B be a symmetric operator such that A ⊂ B. Then, by Theorem 4.11.9, B∗ ⊂ A∗ . Hence A ⊂ B ⊂ B∗ ⊂ A∗ .




Corollary 4.11.16. If A is symmetric and D(A) = H, then A is self-adjoint. Proof: A ⊂ A∗ and D(A) = H implies A∗ = A.




Now we are going to consider the concept of a closed operator. It can be thought of as a weaker version of boundedness. Recall that by the graph of an operator A, defined on D (A) ⊂ E1 and with the range R(A) ⊂ E2 , we mean the set G (A) = {(x, Ax): x ∈ D (A)}. The graph

208

Chapter 4

Linear Operators on Hilbert Spaces

of A : D (A) → R(A) is thus a subset of E1 × E2 . Under our usual assumptions (A is linear and D (A) is a vector space) G (A) is a vector space. Note that, if A ⊂ B, then G (A) ⊂ G (B).

Definition 4.11.17.

(Closed operator) An operator A from a normed space E1 into a normed space E2 is called closed if its graph G (A) is a closed subspace of E1 × E2 , that is, xn ∈ D (A), xn → x, and Axn → y

implies

x ∈ D (A) and Ax = y.

The domain D (A) of a closed operator A need not be closed. It can be proved, that every closed operator from a Banach space into a Banach space is bounded (this is known as the closed graph theorem). Thus, the domain of an unbounded closed operator in a Hilbert space H cannot be closed, and thus it cannot be the entire space H.

Theorem 4.11.18. If A is closed and invertible, then A−1 is closed. Proof: If G (A) = {(x, Ax): x ∈ D(A)} is closed, then, obviously,   
G A−1 = (Ax, x): x ∈ D(A) is closed.




Theorem 4.11.19. If A is densely defined, then A∗ is closed. Proof: If yn ∈ D(A∗ ), yn → y, and A∗ yn → z, then for any x ∈ D(A), we have

$ % Ax, y = lim Ax, yn  = lim x, A∗ yn = x, z. n→∞

n→∞

Hence y ∈ D (A∗ ) and A∗ y = z.




The property of being a closed operator is a desirable one. If a given operator A is not closed, is it possible to extend A to a closed operator ? A possible approach to this problem is to use the closure of G (A) in H × H to define an operator. If cl G (A) defines an operator, then it is a closed extension of A.

Theorem 4.11.20. Let A be an operator in a Hilbert space H. Then there exists

an operator B such that G (B) = cl G (A) if and only if the following condition is satisfied: xn ∈ D (A), xn → 0, Axn → y

implies y = 0.

(4.52)

4.11 Unbounded Operators

209

Proof: Suppose G (B) = cl G (A) for some operator B. If xn ∈ D(A), xn → 0, and Axn → y, then (0, y) ∈ G (B), because G (B) is a closed set. Since (0, 0) ∈ G (B), we must have y = 0. Now assume that (4.52) is satisfied and let (x, y1 ), (x, y2 ) ∈ cl G (A). Then there exist (xn , Axn ), (zn , Azn ) ∈ G (A) such that xn → x, zn → x, Axn → y1 , and Azn → y2 . Since xn − zn ∈ G (A), xn − zn → 0, and A(xn − zn ) → y1 − y2 , (4.52) implies y1 − y2 = 0. This shows that B can be defined by B(x) = y

if (x, y) ∈ cl G (A).

Proof of linearity of B is left as an exercise.




Note that the operator B described in Theorem 4.11.20 is the minimal closed extension of A.

Theorem 4.11.21. Every symmetric, densely defined operator has a closed symmetric extension.

Proof: Let A be a symmetric, densely defined operator in a Hilbert space H.

First we will show that condition (4.52) is satisfied. Let xn ∈ D (A), xn → 0, and Axn → y. Since A is symmetric, we have y, z = lim Axn , z = lim xn , Az = 0, n→0

n→0

for any z ∈ D (A). This implies y = 0, because D (A) is dense in H. Now, by Theorem 4.11.20, there exists a closed operator B such that G (B) = cl G (A), and hence A ⊂ B. It remains to prove that B is symmetric. If x, y ∈ D (B), then there exist xn , yn ∈ D (A) such that xn → x, Axn → Bx, yn → y,

Ayn → By.

Since A is symmetric, we have Axn , yn  = xn , Ayn . By letting n → ∞, we get Bx, y = x, By. Therefore, B is symmetric.




The following is an example of a theorem on solvability of linear equations involving closed densely defined operators. Note the role of these two conditions in the proof of the theorem.

Theorem 4.11.22. Let A be a densely defined closed operator in a Hilbert space H.

210

Chapter 4

Linear Operators on Hilbert Spaces

(a) For any u, v ∈ H, there exist unique x ∈ D (A) and y ∈ D (A∗ ) such that Ax + y = u and x − A∗ y = v. (b) For any v ∈ H, there exists a unique x ∈ D (A∗ A) such that A∗ Ax + x = v.

Proof: (a) Consider the Hilbert space H1 = H × H. Since A is closed, G (A) is a closed subspace of H1 . Thus, by Theorem 3.6.6, H1 = G (A) ⊕ G (A)⊥ , with G (A) ∩ G (A)⊥ = {0}. Now, (z, y) ∈ G (A)⊥ if and only if (x, Ax), (z, y) = 0 for all x ∈ D (A) or, equivalently, x, z + Ax, y = 0 for all x ∈ D (A). Thus, (z, y) ∈ G (A)⊥

if and only if Ax, y = x, −z for all x ∈ D (A).

In other words, (z, y) ∈ G (A)⊥

  if and only if y ∈ D A∗ and z = −A∗ y.

Consequently, if (v, u) ∈ H × H, then there exist unique x ∈ D (A) and y ∈ D(A∗ ) such that   (v, u) = (x, Ax) + −A∗ y, y , which completes the proof of (a). (b) If u = 0 in (a), then it follows that there are unique x ∈ D (A) and y ∈ D (A∗ ) such that Ax + y = 0 and x − A∗ y = v. Thus, x − A∗ (−Ax) = v or A∗ (Ax) + x = v.
We close this section with an interesting observation. Let A be a closed operator in a Hilbert space H. We know that this does not imply boundedness of A. On the other hand, it is always possible to redefine the inner product on D (A) in such a way that D (A) becomes a Hilbert space and A becomes a bounded operator on D (A). In fact, for x, y ∈ D (A) define x, y1 = x, y + Ax, Ay, where ·, · denotes the inner product in H. The proof of completeness of D(A) with respect to the norm x 1 = x 2 + Ax 2 , and the boundedness of A in this new Hilbert space is left as an exercise.

4.12 Exercises

211

4.12 Exercises 1. Let E = {f ∈ L2 (R): f ∈ L2 (R)}. Show that the differential operator is not bounded in E. 2. If A is an operator on a complex Hilbert space H such that Ax ⊥ x for every x ∈ H, show that A = 0. 3. Give an example of a bounded operator A such that A2 = A 2 . 4. Let A be a bounded operator defined on a proper subspace of a Hilbert space H. (a) Define an operator A1 on the closure cl D (A) of the domain of A by A1 x = limn→∞ Axn , where xn ∈ D (A) and xn → x. Show that A1 is well defined, that is, A1 x does not depend on a particular choice of the sequence (xn ). Show that A1 is a linear and bounded operator defined on cl D (A). (b) Define an operator B on H by Bx = A1 x1 , where x1 is the projection of x onto cl D (A). Show that B is a bounded operator on H. (c) Show that A = B . Since A = B on D (A), B is an extension of A. 5. Let ϕ be a symmetric, positive, bilinear functional on a vector space E. Show that   ϕ(x, y)2 ≤ ϕ(x, x)ϕ(y, y). 6. Let (en ) be a complete orthonormal sequence in a Hilbert space H and let (λn ) be a sequence of scalars. (a) Show that there exists a unique operator T on H such that Ten = λn e n . (b) Show that T is bounded if and only if the sequence (λn ) is bounded. (c) For a bounded sequence (λn ), find the norm of T. 7. Let A : R2 → R2 be defined by A(x, y) = (x +2y, 3x +2y). Find the eigenvalues and eigenvectors of A. 8. Let T : R2 → R2 be defined by T(x, y) = (x + 3y, 2x + y). Show that T ∗ = T. 9. Let A : R3 → R3 be given by A(x, y, z) = (3x − z, 2y, −x + 3z). Show that A is self-adjoint. 10. If A is a self-adjoint operator and B is a bounded operator, show that B∗ AB is self-adjoint.

212

Chapter 4

Linear Operators on Hilbert Spaces

11. Prove that the representation T = A + iB in Theorem 4.4.13 is unique. 12. If A∗ A + B∗ B = 0, show that A = B = 0. 13. Let A be an operator on H. Show the following: (a) A is anti-Hermitian if and only if iA is self-adjoint. (b) A − A∗ is anti-Hermitian. 14. If T is self-adjoint and T = 0, show that T n = 0 for all n ∈ N. 15. Let A be a self-adjoint operator. Show the following: (a) Ax + ix 2 = Ax 2 + x 2 . (b) The operator U = (A − iI )(A + iI )−1 is unitary. (U is called the Cayley transform of A after the name of Arthur Cayley (1821–1895).) 16. Show that the limit of a convergent sequence of self-adjoint operators is a self-adjoint operator. 17. If T is a bounded operator on H with one-dimensional range, show that there exist vectors y, z ∈ H such that Tx = x, zy for all x ∈ H. Hence, show the following: (a) T ∗ x = x, yz for all x ∈ H. (b) T 2 = λT, λ is a scalar. (c) T = y z . (d) T ∗ = T if and only if y = αz for some real scalar α. 18. Let A be a bounded self-adjoint operator on a Hilbert space H such that A < 1. Prove that x, (I − A)x ≥ (1 − A ) x 2 for all x ∈ H. 19. Show that the product of isometric operators is an isometric operator. 20. Let (en ) be a complete orthonormal sequence in a Hilbert space H. Show that a bounded operator A on H is unitary if and only if (Aen ) is a complete orthonormal sequence in H. 21. Let (en ), n ∈ Z = {. . . , −2, −1, 0, 1, 2, . . .}, be a complete orthonormal system in a Hilbert space H. Show that there exists a unique bounded operator A on H such that Axn = xn+1 for all n ∈ Z. Operator A is called a two-sided shift operator. Show that A is isometric and unitary. 22. Show that the product of two unitary operators is a unitary operator. 23. Let A be a bounded operator on a Hilbert space. Define the exponential operator by

4.12 Exercises

e = A

∞  An n=0

n!

213

  0 A =I .

Show that eA is a well-defined operator. Prove the following: (a) (eA )n = enA for any n ∈ N. (b) e0 = I . (c) eA is invertible (even if A is not) and its inverse is e−A . (d) eA eB = eA+B for any commuting operators A and B. (e) If A is self-adjoint, then eiA is unitary. 24. If T is a normal operator on H and λ is a scalar, show that  ∗  T x − λx = Tx − λx for all x ∈ H. 25. Show that, if the kernel K(x, y) satisfies K(x, y) = K(y, x), then for any real α the operator 

b

(Tu)(x) = αu(x) + i

K(x, y)u(y)dy a

on L2 ([a, b]) is normal. 26. Show that, for any invertible operator T, the operator T ∗ T is also invertible. 27. Let T and S be commuting operators. Show that if both T and S are normal, then S + T and ST are normal. 28. If T ∗ T = I , is it true that TT ∗ = I ? 29. Let A, B, C, and D be positive operators on a Hilbert space. Prove the following: (a) If A ≥ B and C ≥ D, then A + C ≥ B + D. (b) If A ≥ 0 and α ≥ 0 (α ∈ R), then αA ≥ 0. (c) If A ≥ B and B ≥ C, then A ≥ C. 30. If A is a positive operator and B is a bounded operator, show that B∗ AB is positive. 31. If A and B are positive operators and A + B = 0, show that A = B = 0. 32. Show that for any self-adjoint operator A, there exists positive operators S and T such that A = S − T and ST = 0.

214

Chapter 4

Linear Operators on Hilbert Spaces

33. Show that if A is a positive definite operator, then it is invertible and its inverse is positive definite. 34. Find all operators T : R2 → R2 such that T 2 = I . Which one is the positive square root of I ? 35. Find the positive square root of the operator T on L2 ([a, b]) defined by (Tf )(t) = g(t)f (t), where g is a positive continuous function on [a, b]. √ √ 36. Show that A = A . 37. Let A and B be positive operators on a Hilbert space. Show that A2 = B2 implies A = B. √ 38. √ Let √ A and B be commuting positive operators. Show that AB = A B. 39. If P is self-adjoint and P2 is a projection operator, is P a projection operator ? 40. Let T be a multiplication operator on L2 ([a, b]). Find necessary and sufficient conditions for T to be a projection. 41. Give an example of two noncommuting projection operators. 42. Show that P is a projection if and only if P = P∗ P. 43. If P, Q, and P + Q are projections, show that PQ = 0. 44. Show that every projection P is a positive operator and 0 ≤ P ≤ I . 45. If T is an isometric operator, show that TT ∗ is a projection. 46. Show that for projections P and Q, the operator P + Q − PQ is a projection if and only if PQ = QP. 47. Prove that the collection of all compact operators on a Hilbert space H is a vector space. 48. Show that the projection onto a closed subspace F of a Hilbert space H is a compact operator if and only if F is finite-dimensional. 49. Show that the operator T : l2 → l2 defined by T(xn ) = (2−n xn ) is compact. 50. Show that a self-adjoint operator T is compact if and only if there exists a sequence of finite-dimensional operators strongly convergent to T. 51. Show that the space of all eigenvectors corresponding to a nonzero eigenvalue of a compact operator is finite-dimensional. 52. Show that eigenvalues of a symmetric operator are real and eigenvectors corresponding to different eigenvalues are orthogonal.

4.12 Exercises

215

53. Show that shift operators have no eigenvalues. 54. Give an example of a self-adjoint operator which has no eigenvalues. 55. Give an example of a normal operator which has no eigenvalues. 56. Show that a nonzero vector x is an eigenvector of an operator A if and only if |Ax, x| = Ax x . 57. Show that if the eigenvectors of a self-adjoint operator A form a complete orthogonal system and all eigenvalues are non-negative (or positive), then A is positive (or strictly positive). 58. Prove the spectral theorem for the finite-dimensional case: If T : RN → RN is a self-adjoint operator, then there exists an orthonormal system of vectors ϕ1 , . . . , ϕN ∈ RN and scalars λ1 , . . . , λN ∈ C such that Tϕk = λk ϕk , k = 1, . . . , N. Hence, the matrix corresponding to T relative to the basis (ϕ1 , . . . , ϕN } is   λ1 0 . . . 0  0 λ2 . . . 0   . . .. . . ..   .. . . .  0

0

...

λN

59. If λ is an approximate eigenvalue of an operator T, show that |λ| ≤ T . 60. Show that if T has an approximate eigenvalue λ such that |λ| = T , then sup x ≤1 |Tx, x| = T . 61. Show that for every bounded self-adjoint operator T on a Hilbert space, at least one of the numbers T or − T is an approximate eigenvalue. 62. If λ is an approximate eigenvalue of T, show that λ + µ is an approximate eigenvalue of T + µI and λµ is an approximate eigenvalue of µT. 63. Show that for every approximate eigenvalue λ of an isometric operator, we have |λ| = 1. 64. Show that every approximate eigenvalue of a self-adjoint operator is real. 65. Show that if λ is an approximate eigenvalue of a normal operator T, then λ is an approximate eigenvalue of T ∗ . 66. For arbitrary unbounded operators A, B, and C, prove the following: (a) (A + B)C = AC + BC. (b) AB + AC ⊂ A(B + C). Give an example of operators A, B, and C for which AB+AC = A(B+C). 67. Show that (A + B)∗ ⊃ A∗ + B∗ .

216

Chapter 4

Linear Operators on Hilbert Spaces

68. Give an example of a closed operator whose domain is not a closed set. 69. Show that A∗∗ is symmetric whenever A is symmetric. 70. If A is an operator on a Hilbert space H and there exists an operator B on H such that Ax, y = x, By for all x, y ∈ H, show that A is bounded and B = A∗ . 71. Let A be a closed operator in a Hilbert space H. Prove that D (A) is a Hilbert space with respect to the inner product defined by x, y1 = x, y + Ax, Ay, where ·, · denotes the inner product in H. Prove that A is a bounded operator on D (A) with the defined inner product.

Applications to Integral and Differential Equations

CHAPTER

5

“Since a general solution must be judged impossible from want of analysis, we must be content with the knowledge of some special cases, and that all the more, since the development of various cases seems to be the only way to bringing us at last to a more perfect knowledge.” Leonhard Euler “The profound study of nature is the most fertile source of mathematical discoveries.” Joseph Fourier

5.1 Introduction In this chapter we discuss some applications of the theory of Hilbert spaces to integral and differential equations. The goal is to illustrate possibilities of applications of techniques developed in Chapters 3 and 4. In Section 5.2, we give a number of important existence and uniqueness theorems for operator equations. In subsequent sections, particular applications are discussed. More specifically, Sections 5.3 to 5.7 deal with different kinds of integral equations and their applications. Ordinary differential equations and differential operators are discussed in Section 5.8. Section 5.9 is devoted to Sturm–Liouville systems. Section 5.10 deals with inverse differential operators and Green’s functions. Section 5.11 is devoted to an important operator on L2 (R): the Fourier transform. Applications of the Fourier transform to ordinary and integral equations are discussed in Section 5.12.

217

218

Chapter 5

Integral and Differential Equations

Equations with the unknown function under the integral sign are called integral equations. The equation of the form  x ϕ(x) = K(x, t)f (t)dt, a

where ϕ is given and f is the unknown function, is called a Volterra equation of the first kind (Vito Volterra (1860–1940)). The function K is called the kernel. If the unknown function appears also outside the integral, that is,  x f (x) = K(x, t)f (t)dt + ϕ(x), a

then the equation is called a Volterra equation of the second kind. Similarly, the equations with fixed limits of integration 

b

ϕ(x) =

K(x, t)f (t)dt a

and



b

f (x) =

K(x, t)f (t)dt + ϕ(x),

a

are called Fredholm equations of the first and second kind, respectively (Erik Ivar Fredholm (1866–1927)). If ϕ = 0, the equation is called homogeneous, otherwise it is called nonhomogeneous.

5.2 Basic Existence Theorems Equations in applied mathematics can often be written as operator equations of the form Tx = x,

(5.1)

where T is an operator in a Hilbert space and x is the unknown. It is easy to recognize that (5.1) is an eigenvalue equation with the eigenvalue λ = 1. As in Section 5.6, solutions of Equation (5.1) will be called fixed points of the mapping T. Fixed points are thus elements of the space that are unchanged by the action of T. There is a large body of mathematics dealing with the existence and determination of fixed points. Recall from Chapter 1 (see Definition 1.6.2) that by a contraction mapping we mean a mapping T : E → E, where E is a subset of a normed space, for which there exists a positive number α < 1 such that Tx − Ty ≤ α x − y

for all x, y ∈ E.

(5.2)

5.2 Basic Existence Theorems

219

Now we are going to use the Banach fixed point theorem (Theorem 1.6.4) to obtain the existence and uniqueness of solutions of certain operator equations. For convenience, we restate the theorem in a slightly different form geared for applications:

Theorem 5.2.1. (Contraction mapping theorem) Let S be a closed subset of a Banach space, and let T : S → S be a contraction mapping. Then (a) the equation Tx = x has one and only one solution in S, and (b) the unique solution x can be obtained as the limit of the sequence (xn ) of elements of S defined by xn = Txn−1 , n = 1, 2, . . . , where x0 is an arbitrary element of S: x = lim T n x0 . n→∞

(5.3)

This theorem not only represents an existence and uniqueness result but also gives an algorithm for finding the solution by an iterative procedure, which is known as the method of successive approximations. Theorem 5.2.1 can be used to prove existence and to find solutions of algebraic, differential, and integral equations. The following is a useful generalization of Theorem 5.2.1. It will be essential in the proof of the existence and uniqueness of solutions of a certain type of integral equations; see Theorem 5.5.1.

Theorem 5.2.2. Let E be a Banach space, and let T : E → E. If T m is a contraction

for some m ∈ N, then T has a unique fixed point x0 ∈ E and x0 = limn→∞ T n x for any x ∈ E.

Proof: In view of Theorem 5.2.1, T m has a unique fixed point x0 ∈ E. Then

T m x0 = x0 and hence

  T m (Tx0 ) = T T m x0 = Tx0 .

By uniqueness of the fixed point of T m , we must have Tx0 = x0 . The proof of the fact that x0 = limn→∞ T n x for any x ∈ E is left as an exercise.


Theorem 5.2.3. If A is a bounded linear operator on a Banach space E, and ϕ is an arbitrary element of E, then the operator defined by Tf = αAf + ϕ

(5.4)

has a unique fixed point for any sufficiently small |α|. More precisely, if k is a positive constant such that Af ≤ k f for all f ∈ E, then Tf = f has a unique solution whenever |α|k < 1.

220

Chapter 5

Integral and Differential Equations

Proof: Since A is bounded, there exists a constant k such that Af1 − Af2 ≤ k f1 − f2

for all f1 , f2 ∈ E.

Thus, Tf1 − Tf2 = |α| Af1 − Af2 ≤ |α|k f1 − f2 , and hence T is a contraction whenever |α| < 1/k. In such a case T has a unique fixed point by Theorem 5.2.1.
When the iterative process is applied in the case described in the previous theorem, we obtain the following sequence approximating the solution: f0 = an arbitrary element of E, f1 = Tf0 = αAf0 + ϕ, f2 = T(αAf0 + ϕ) = α 2 A2 f0 + αAϕ + ϕ, .. . fn = α n An f0 + α n−1 An−1 ϕ + · · · + α 2 A2 ϕ + αAϕ + ϕ, .. . . Therefore, the solution f can be written as f = ϕ + αAϕ + α 2 A2 ϕ + · · · + α n An ϕ + · · · .

(5.5)

Note that the choice of f0 is irrelevant for the solution. However, since some choices of f0 give faster convergence of the series, in applications it may be important to make a good “first guess.” Formally, the expansion (5.5) can be obtained directly from the equation f − αAf = ϕ by expanding (I − αA)−1 into a geometric series (I − αA)−1 = I + αA + α 2 A2 + · · · .

(5.6)

Theorem 4.9.8 provides a rigorous justification for this equality.

Corollary 5.2.4. Let A be a bounded linear operator in a Banach space. Then the equation x = x0 + αAx has a unique solution given by x=

∞  n=0

α n An x 0

5.2 Basic Existence Theorems

221

whenever |α| A < 1. The following theorem, attributed to Charles Emile Picard (1856–1941), is an important example of the application of the fixed point theorem to ordinary differential equations.

Theorem 5.2.5. (Picard’s existence and uniqueness theorem) Consider the initial value problem for the ordinary differential equation dy = f (x, y) dx

(5.7)

y(x0 ) = y0 ,

(5.8)

with the initial condition where f is a continuous function in some closed domain
 R = (x, y): a ≤ x ≤ b, c ≤ y ≤ d containing the point (x0 , y0 ) in its interior. If f satisfies the Lipschitz condition   f (x, y1 ) − f (x, y2 ) ≤ K|y1 − y2 | (5.9) for some K ∈ R and all (x, y1 ), (x, y2 ) ∈ R, then there exists a unique solution y = ϕ(x) of the problem (5.7) to (5.8) defined in some neighborhood of x0 .

Proof: Observe that every solution of the integral equation  y(x) = y0 +

x

  f t, y(t) dt

(5.10)

x0

satisfies (5.7) and (5.8), and conversely. Consider the operator T defined on C ([a, b]) by  x   (Tϕ)(x) = y0 + f t, ϕ(t) dt. (5.11) x0

Let

 
 M = sup f (x, y): (x, y) ∈ R ,

and select ε > 0 such that Kε < 1 and [x0 − ε, x0 + ε] ⊂ [a, b]. If    
S = ϕ(x) ∈ C [x0 − ε, x0 + ε] : ϕ(x) − y0  ≤ Mε  for all x ∈ [x0 − ε, x0 + ε] , then S is a closed subset of the Banach space C ([x0 − ε, x0 + ε]) with the sup-norm   ϕ = sup ϕ(x). [x0 −ε,x0 +ε]

222

Chapter 5

Integral and Differential Equations

Furthermore, if ϕ ∈ S and x ∈ [x0 − ε, x0 + ε], then     x     (Tϕ)(x) − y0  =  f t, ϕ(t) dt  ≤ Mε,  x0

and thus T maps S onto itself. Finally, for any ϕ1 , ϕ2 ∈ S, we have   x         f t, ϕ1 (t) − f t, ϕ2 (t) dt  ≤ Kε ϕ1 − ϕ2 . Tϕ1 − Tϕ2 = sup  [x0 −ε,x0 +ε]

x0

Thus, since Kε < 1, T is a contraction. Therefore, in view of Theorem 5.2.1, there is a unique solution ϕ of the equation Tϕ = ϕ, that is, y = ϕ is a unique solution of (5.10).
We close this section with a version of the so-called Fredholm alternative. It is one of the most important theorems in applied mathematics. It gives a definite criterion for existence of solutions of linear operator equations. Roughly speaking, it replaces the question of uniqueness of solution of a nonhomogeneous equation by the question of existence of a nontrivial solution of the corresponding homogeneous equation.

Theorem 5.2.6. (Fredholm alternative for self-adjoint compact operators) Let A be a self-adjoint compact operator on a Hilbert space H. Then the nonhomogeneous operator equation f = Af + ϕ

(5.12)

has a unique solution for every ϕ ∈ H if and only if the homogeneous equation g = Ag

(5.13)

has only trivial solution g = 0. Moreover, if Equation (5.12) has a solution, then ϕ, g = 0 for every solution g of (5.13).

Proof: In view of Corollary 4.10.2, H has an orthonormal basis (vn ) consisting of eigenvectors of A with corresponding eigenvalues (λn ). Let ∞ 

ϕ=

cn vn .

(5.14)

n=1

We seek a solution of (5.12) in the form f=

∞ 

an vn .

n=1

Thus, ∞  n=1

an vn =

∞  n=1

an λn vn +

∞  n=1

cn vn ,

223

5.2 Basic Existence Theorems

and hence, an =

cn 1 − λn

for all n ∈ N,

(5.15)

provided λn = 1. If (5.13) has no nonzero solutions, 1 is not an eigenvalue of A, and hence (5.15) is valid. Therefore, if (5.12) has a solution, it has to be of the form f=

∞  n=1

cn vn . 1 − λn

(5.16)

This shows that if (5.12) has a solution, then it is unique. To prove that (5.12) has a solution, it suffices to show that the series (5.16) is always convergent. By Theorem 4.9.19, we have λn → 0, and therefore, 1 ≤M 1 − λn for some constant M and all n ∈ N. Consequently, ∞    cn  1 − λ n=1

n

2 ∞    ≤ M2 |cn |2 < ∞.  n=1

Thus, by Theorem 3.4.10, series (5.16) converges and its sum is a solution of (5.12). If (5.13) has a nontrivial solution g and f is a solution of (5.12), then also f + cg is a solution of (5.12) for any c ∈ C. Therefore, (5.12) has infinitely many solutions in that case. Finally, suppose f and g are solutions of (5.12) and (5.13), respectively. Then f , g = Af , g + ϕ, g = f , Ag + ϕ, g = f , g + ϕ, g, since Ag = g. This means ϕ, g = 0. Hence, if (5.12) has a solution, then ϕ is orthogonal to every solution of (5.13).
Theorem 5.2.6 is called an alternative because of the following version: Either the equation f = Af + ϕ has a solution for every ϕ or the corresponding homogeneous equation f = Af has a nontrivial solution.

224

Chapter 5

Integral and Differential Equations

5.3 Fredholm Integral Equations In this and the following sections, we study solvability of integral equations.

Theorem 5.3.1. (Existence and uniqueness of solution of the Fredholm nonhomogeneous linear integral equation of the second kind) The equation 

b

f (x) = α

K(x, y)f (y)dy + ϕ(x)

(5.17)

a

has a unique solution f ∈ L2 ([a, b]) provided the kernel K is continuous in [a, b] × [a, b], ϕ ∈ L2 ([a, b]), and |α|k < 1, where +  b b   K(x, y)2 dxdy. k= a

a

Proof: Consider the operator 

b

(Tf )(x) = α

K(x, y)f (y)dy + ϕ(x).

a

Since ϕ ∈ L2 ([a, b]), Tf ∈ L2 ([a, b]) if 

b

  K(x, y)f (y)dy ∈ L2 [a, b] .

(5.18)

a

By Schwarz’s inequality, we find   b  b      K(x, y)f (y)dy K(x, y)f (y)dy ≤  a

a



b

 K(x, y)2 dy

≤ a

1/2 

b

 f (y)2 dy

1/2 .

a

Therefore, 2  b  b  b    2  2        K(x, y) dy f (y) dy K(x, y)f (y)dy ≤  a

a

a

and 2   b  b  b  b  b    2  2       K(x, y) dy f (y) dy dx K(x, y)f (y)dy dx ≤  a

a

a

≤

 b a

a

a

b

 K(x, y)2 dydx

a

 a

b

 f (y)2 dy.

5.3 Fredholm Integral Equations

225

Since  b a

b

 K(x, y)2 dydx < ∞ and

a



b

 f (y)2 dy < ∞,

a

(5.18) is satisfied and thus, T maps L2 ([a, b]) into itself. Note that the above shows also that the operator defined by 

b

(Af )(x) =

K(x, y)f (y)dy a

is bounded. Therefore, by Theorem 5.2.3, the equation Tf = f has a unique solution whenever |α|k < 1.


Example 5.3.2. Consider the integral equation  f (x) = α

b

e(x−y)/2 f (y)dy + ϕ(x),

(5.19)

a

where ϕ is a given function. Since  b a

b a

2 (eb − ea )2 , e(x−y)/2 dxdy = ea+b

Equation (5.19) has a unique solution whenever |α|
2.

a

The kernel can be also written as  b  b ... K(x, ξn−1 )K(ξn−1 , ξn−2 ) · · · K(ξ1 , t)dξn−1 dξn−2 . . . dξ1 . Kn (x, t) = a

a

We next apply Corollary 5.2.4 to obtain the following result regarding the solvability of (5.21) and hence also the integral equation (5.22). If |α| T < 1, then Equation (5.21) has a unique solution given by the Neumann series ∞  f =ϕ+ α n T n ϕ. (5.23) n=1

Hence, the integral equation (5.22) has a unique solution f given by '  b & ∞ α n−1 Kn (x, t) ϕ(t)dt. (5.24) f (x) = ϕ(x) + α a

n=1

If we adopt the notation (x, t; α) =

∞ 

α n−1 Kn (x, t),

n=1

then the solution can be written in the form  b (x, t; α)ϕ(t)dt. f (x) = ϕ(x) + α a

(5.25)

228

Chapter 5

Integral and Differential Equations

The function  is often called the resolvent kernel.

Example 5.4.1. We obtain the Neumann series solution of the integral equation 1 f (x) = x + 2



1 −1

(t − x)f (t)dt.

First we set f0 (x) = x. Then 1 f1 (x) = x + 2



1

1 (t − x)tdt = x + . 3 −1

Substituting f1 back into the original equation, we find    1 1 1 1 1 dt = x + − x. f1 (x) = x + (t − x) t + 2 −1 3 3 3 Continuing this process, we obtain 1 x 1 − − 2, 3 3 3 1 x x 1 f4 (x) = x + − − 2 + 2 , 3 3 3 3 .. . n n   m−1 −m (−1) 3 −x (−1)m−1 3−m . f2n (x) = x + f3 (x) = x +

m=1

m=1

By letting n → ∞, we get 3 1 f (x) = x + . 4 4 It is easy to verify this solution by direct substitution into the original equation.


5.5 Volterra Integral Equations The Volterra equations 

x

K(x, y)f (y)dy = ϕ(x),

(5.26)

a

and



x

f (x) − α a

K(x, y)f (y)dy = ϕ(x)

(5.27)

5.5 Volterra Integral Equations

229

are special cases of the Fredholm equations. The latter reduces to the corresponding Volterra equation if K(x, y) = 0 for y > x. However, Volterra equations have many important and interesting properties that cannot be obtained from the theory of Fredholm equations, so a separate study is definitely needed.

Theorem 5.5.1. (Volterra equation of the second kind) Suppose ϕ ∈ L2 ([a, b]) and the kernel K satisfies the condition  b a

b

 K(x, y)2 dxdy < ∞.

(5.28)

a

Then the equation 

x

f (x) = α

K(x, y)f (y)dy + ϕ(x)

(5.29)

a

has a unique solution in L2 ([a, b]) for arbitrary α ∈ C. The solution can be written in the form f (x) = ϕ(x) +

∞ 



x

αn

Kn (x, t)ϕ(t)dt,

(5.30)

a

n=1

where the kernels Kn (x, t) satisfy the recurrence relation K1 (x, t) = K(x, t),  x K(x, ξ )Kn−1 (ξ, t)dξ, Kn (x, t) =

for n ≥ 2.

(5.31)

a

Proof: We set  A(x) =

x

 K(x, y)2 dy

 and

b

 K(x, y)2 dx.

B(y) =

a

y

By (5.28), A and B are integrable functions, so that there exists a constant M such that  b  b A(x)dx ≤ M and B(y)dy ≤ M. a

a

We also introduce the function λ on [a, b] defined by  x λ(x) = A(t)dt. a

Clearly, 0 ≤ λ(x) ≤ M for all x ∈ [a, b].

230

Chapter 5

Integral and Differential Equations

Consider the operator 

x

(Tf )(x) = α

K(x, y)f (y)dy + ϕ(x).

a

We shall show that T n is a contraction for some n ∈ N, and then use Theorem 5.2.2 to conclude that T has a fixed point. That fixed point must be a unique solution of (5.29). If we write Tf = αWf + ϕ, where



x

(Wf )(x) =

K(x, y)f (y)dy, a

then T n f = ϕ + αWϕ + α 2 W 2 ϕ + · · · + α n W n f . The operators W m can be written in the form  x  m  Km (x, y)g(y)dy, W g (x) = a

where the kernels Kn are defined by (5.31). Indeed, for m = 2 we have  x  z  2  K(x, z) K(z, y)g(y)dydz. W g (x) = a

a

This integral can be considered as a double integral over the triangular region {(y, z): a ≤ y ≤ z and a ≤ z ≤ x} (see Figure 5.1). After interchanging the order of integration, we obtain  x x  2  W g (x) = K(x, z)K(z, y)dzg(y)dy. a

If we denote

y



x

K2 (x, y) =

K(x, z)K(z, y)dz, y

then by a similar argument we get  x x  3  W g (x) = K(x, z)K2 (z, y)dzg(y)dy, a

and so on, as stated earlier.

y

5.5 Volterra Integral Equations

231

Figure 5.1 The triangular region {(y, z): a ≤ y ≤ z and a ≤ z ≤ x}. To estimate W m , we examine Km . For m = 2, application of Schwarz’s inequality gives  x 2     K2 (x, y)2 =  K(x, z)K(z, y)dz  y



x

 K(x, z)2 dz

≤



y

x

 K1 (z, y)2 dz ≤ A(x)B(y).

y

Similarly,   K3 (x, y)2 ≤



x

 K(x, z)2 dz

y



≤ A(x)B(y)



x

 K2 (z, y)2 dz

y x

  A(z)dz = A(x)B(y) λ(x) − λ(y) .

y

By induction, we can show that m−2   Km (x, y)2 ≤ A(x)B(y) (λ(x) − λ(y)) (m − 2)!

for m ≥ 2.

Therefore,   x  m     2 T f1 (x) − T m f2 (x)2 = |α|2m  Km (x, y) f1 (y) − f2 (y) dy  a  x  x   A(x)B(y)[λ(x) − λ(y)]m−2 2m f1 (y) − f2 (y)2 dy dy ≤ |α| (m − 2)! a a 2m m−2  x |α| A(x)(λ(x)) ≤ B(y)dy f1 − f2 2 (m − 2)! a ≤

|α|2m A(x)(λ(x))m−2 M f1 − f2 2 . (m − 2)!

232

Chapter 5

Integral and Differential Equations

Integrating with respect to x in [a, b], we obtain 2m m   m T f1 − T m f2 2 ≤ |α| M f1 − f2 2 (m − 1)!

for m ≥ 2.

Therefore, since there exists n ∈ N such that |α|2n M n < 1, (n − 1)! T n is a contraction. In view of Theorem 5.2.2, Equation (5.29) has a unique solution that can be written in the form lim T n f = ϕ + αWϕ + α 2 W 2 ϕ + α 3 W 3 ϕ + · · · ,

n→∞

or, equivalently, f (x) = ϕ(x) +

∞ 



x

αn

Kn (x, t)ϕ(t)dt.




a

n=1

Note that the restriction on α imposed in the result of Section 5.4 is not needed in Theorem 5.5.1.

Theorem 5.5.2. (Homogeneous Volterra equation) The homogeneous Volterra equation



x

f (x) = α

K(x, t)f (t)dt,

x ∈ [0, 1]

(5.32)

0

has only the trivial solution f = 0.

Proof: From (5.32), we have   f (x) ≤ |α|



x

  K(x, t)f (t)dt ≤ |α|Mp,

(5.33)

0

where



1

 f (t)dt

p= 0

and M is a constant such that |K(x, t)| ≤ M for all x, t ∈ [0, 1]. Hence, by using (5.33) in (5.32), we obtain  x     f (x) ≤ |α| K(x, t)|α|Mpdt ≤ |α|2 M 2 px. 0

By continuing the process, we get n−1   |α|n M n p f (x) ≤ |α|n M n p x ≤ → 0 as n → ∞. (n − 1)! (n − 1)!

5.6 Method of Solution for a Separable Kernel

This shows f (x) = 0 for all x ∈ [0, 1].

233


5.6 Method of Solution for a Separable Kernel It is of interest to examine the solvability of the Fredholm integral equation of the second kind with separable kernel. The kernel K is called separable if it is of the form n  K(s, t) = Mk (s)Nk (t), (5.34) k=1

where the functions Mk and Nk are assumed to belong to L2 ([a, b]), so that the kernel is square integrable. Such kernels are also called degenerate. They include all polynomials and many transcendental functions, for instance cos(s + t) = cos s cos t − sin s sin t. For separable kernels, the Fredholm equation of the second kind can be put into the form  b n  f (x) = ϕ(x) + α Mk (x) Nk (t)f (t)dt (5.35) a

k=1

or f (x) = ϕ(x) + α

n 

ck Mk (x),

(5.36)

k = 1, . . . , n.

(5.37)

k=1

where



b

ck =

Nk (t)f (t)dt, a

The constants ck depend on the unknown solution f and therefore, they are not known. On the other hand, we know that the solution of Equation (5.35) is of the form (5.36), and thus the problem is reduced to determination of the constants c1 , . . . , cn . This can be done in the following way. First, we multiply Equation (5.36) by Nm (x) and integrate with respect to x to eliminate the x-dependence. Consequently, it follows by (5.37) that cm = α

n 

amk ck + bm ,

(5.38)

k=1

where 



b

amk =

Nm (x)Mk (x)dx a

and

b

bm =

Nm (x)ϕ(x)dx. a

234

Chapter 5

Integral and Differential Equations

Equation (5.38) can be written in the matrix form (I − αA)c = b or c = (I − αA)−1 b, where A = (amk ), b = (b1 , . . . , bn ), and c = (c1 , . . . , cn ). This equation is equivalent to a system of linear simultaneous algebraic equations of the form (1 − αa11 )c1 − αa12 c2 −αa21 c1 + (1 − αa22 )c2 .. .

− αa13 c3 − · · · − αa23 c3 − · · · .. .

−αan1 c1 − αan2 c2

− αan3 c3 − · · ·

− αa1n cn = b1 , − αa2n cn = b2 , .. .

+ (1 − αann )cn = bn .

This system has a unique solution if det(I − αA) = 0. If the original equation is homogeneous, that is, if ϕ(x) = 0, then b = 0. Then the system of homogeneous equations (I − αA)c = 0

(5.39)

has a nontrivial solution if and only if det(I − αA) = 0. We can use the roots of this equation to find solutions of (5.35). Note that in this case we have infinitely many solutions.

Example 5.6.1. We use the above method to solve the integral equation 

1

f (x) = α

(1 − 3xt)f (t)dt.

(5.40)

0

We have K(x, t) = 1 − 3xt =

2 

Mk (x)Nk (t),

k=1

where M1 (x) = 1,

N1 (t) = 1,

Since



M2 (x) = −3x,

N2 (t) = t.

1

amk =

Nm (x)Mk (x)dx, 0

we have a11 = 1,

3 a12 = − , 2

1 a21 = , 2

a22 = −1.

5.6 Method of Solution for a Separable Kernel

235

The solution of (5.40) has the form f (x) = α

2 

ck Mk (x),

k=1

where



1

ck =

Nk (t)f (t)dt,

k = 1, 2.

0

The system of algebraic equations associated with this problem is 3 (1 − α)c1 + α c2 = 0, 2 1 −α c1 + (1 + α)c2 = 0. 2 The roots of the equation

 1 − α   −1α 2

  =0 1+α 3 2α

are α = 2 and α = −2. Substituting them into the algebraic system, we obtain c1 = 3c2 c1 = c2

for α = 2, for α = −2.

Hence the eigenfunctions are   for α = 2; f1 (x) = 2 c1 M1 (x) + c2 M2 (x) = 2(c1 − 3c2 x) = a(1 − x)   f2 (x) = −2 c1 M1 (x) + c2 M2 (x) = −2(c1 − 3c2 x) = b(1 − 3x) for α = −2. Therefore, Equation (5.40) has nontrivial solution only if α = 2 or α = −2 and the solutions are f (x) = a(1 − x) and f (x) = b(1 − 3x), respectively; a and b are arbitrary scalars.


Example 5.6.2. Consider now a nonhomogeneous equation 

1

f (x) = ϕ(x) + α

(1 − 3xt)f (t)dt.

0

If

 1 − α   −1α 2

   = 0, 1+α 3 2α

then the equation has a unique solution of the form f (x) = ϕ(x) + α

2  k=1

ck Mk (x),

(5.41)

236

Chapter 5

Integral and Differential Equations

where c1 and c2 are solutions of the linear system 3 (1 − α)c1 + α c2 = b1 = 2 1 −α c1 + (1 + α)c2 = b2 = 2

 

1



0 1

1

N1 (x)ϕ(x)dx =

ϕ(x)dx, 

0 1

N2 (x)ϕ(x)dx =

0

xϕ(x)dx. 0

On the other hand, if α = 2, then the equation has a solution if and only if 

1

ϕ(x)(1 − x)dx = 0,

0

and in this case the solution is given by 

1

f (x) = ϕ(x) − 2

ϕ(ξ )dξ + a(1 − x).

0

Finally, if α = −2, the equation has a solution if and only if 

1

ϕ(x)(1 − 3x)dx = 0,

0

and in this case the solution is  f (x) = ϕ(x) − 2

1

ξ ϕ(ξ )dξ + b(1 − 3x).




0

5.7 Volterra Integral Equations of the First Kind and Abel’s Integral Equation In general, it is difficult to deal with Volterra integral equations of the first kind. In this section, we consider some special cases which can be solved without much difficulty. We consider the equation 

x

K(x, t)f (t)dt = ϕ(x),

x ∈ [0, 1],

(5.42)

0

where K and ϕ are assumed to be differentiable. We differentiate (5.42) with respect to x to obtain  x ∂ K(x, x)f (x) + K(x, t)f (t)dt = ϕ (x). 0 ∂x

5.7 Abel’s Integral Equation

237

If K(x, x) = 0, this equation can be transformed into a Volterra integral equation of the second kind:  x ∂ ϕ (x) ∂x K(x, t) f (t)dt = . (5.43) f (x) + K(x, x) K(x, x) 0 If the kernel and the function on the right-hand side of (5.43) are square integrable, then (5.43) has a unique solution in L2 ([0, 1]) by the general theory of integral equations.

Example 5.7.1. (Abel’s equation) An equation of the form  0

x

f (t) dt = ϕ(x), (x − t)α

(5.44)

where 0 ≤ α < 1, ϕ is continuous and ϕ(0) = 0, is called Abel’s equation. We denote the kernel of (5.44) by Kα . If we apply the Volterra integral operator with kernel Kβ (0 ≤ β < 1) to both sides of (5.44), we obtain  x  x , x dz ϕ(t) dt. (5.45) f (t)dt = β α β 0 t (x − z) (z − t) 0 (x − t) Next we put z = t + (x − t)v inside the square bracket on the left-hand side of (5.45) to obtain  x  1 dz dv 1−α−β = (x − t) β α β α t (x − z) (z − t) 0 (1 − v) v = (x − t)1−α−β

(1 − α)(1 − β) , (2 − α − β)

where  denotes the Euler gamma function (Leonhard Euler (1707–1783)):  ∞ (p) = xp−1 e−x dx, p > 0. 0

Consequently, the integral Equation (5.45) becomes  x  x (2 − α − β) f (t) ϕ(t) dt = dt. α+β−1 (1 − α)(1 − β) 0 (x − t)β 0 (x − t) In particular, if we set a + β − 1 = 0, this takes a simple form  x   x ϕ(t) ϕ(t) (1) sin π α x f (t)dt = dt = dt, 1−α (α)(1 − α) (x − t) π (x − t)1−α 0 0 0 (5.46)

238

Chapter 5

Integral and Differential Equations

because (1) = 1 and (α)(1 − α) = π/ sin π α. If we assume that the right-hand side of (5.46) is differentiable, we can differentiate both sides to obtain  ϕ(t) sin π α d x dt. (5.47) f (x) = π dx 0 (x − t)1−α This is the desired solution of (5.44). In particular, if α = 12 , then Abel’s equation (5.44) has the solution  1 d x ϕ(t) f (x) = dt. √ π dx 0 x−t In 1825, Niels Henrik Abel (1802–1829) solved this equation in connection with the problem of tautochronous motion. This problem deals with the motion of a heavy particle sliding under the action of gravitational acceleration g, without a friction, down a curve which passes through the origin. We want to find a curve for which the particle descends to its lowest point in a constant time independent of the initial position. The velocity of the particle at any point (ξ, η) satisfies the following equation: ds v = = 2g y − η, (5.48) dt where the initial position of the particle is (x, y), s is the arc length of the curve from the initial point (x, y) to (ξ, η), and t is the time. Integrating (5.48), we obtain  T  0  y ds ds 2g dt = − = . √ √ y − η y −η 0 y 0 Assuming s = f (η) with f (0) = 0, we get  y f (η) 2gT = dη. 1/2 0 (y − η)

(5.49)

This is Abel’s equation with α = 12 , and so its solution is 1 d f (y) = π dy



y 0

√ √ √  2gT dη T 2g d y T 2g dη = = √ . (y − η)1/2 π dy 0 (y − η)1/2 π y

Thus, f (y) = 2 2ay,

gT 2 . (5.50) π This plane curve represents a cycloid with vertex at the origin and the x-axis as the tangent at the vertex.
a=

5.8 Ordinary Differential Equations

239

5.8 Ordinary Differential Equations and Differential Operators In this section, we consider operators arising from boundary value problems associated with second-order ordinary differential equations. This type of an ordinary differential equation is most common in applied mathematics, science, and engineering. Consider the second-order linear differential equation of the form Lu = f

(5.51)

over [a, b] subject to two nonhomogeneous boundary conditions B1 (u) = b1 ,

(5.52)

B2 (u) = b2 ,

(5.53)

where B1 and B2 are functions of the unknown solution u and its derivatives, and b1 , b2 are constants. Finally, assume that the operator L has the form (Lu)(x) = a2 (x)

d2 d u(x) + a1 (x) u(x) + a0 (x)u(x), 2 dx dx

(5.54)

where a2 (x) = 0 on [a, b]. In this section, we consider only those functions B1 and B2 for which boundary conditions (5.52) and (5.53) can be written as B1 (u) = α11 u(a) + α12 u (a) + β11 u(b) + β12 u (b) = b1 ,



B2 (u) = α21 u(a) + α22 u (a) + β21 u(b) + β22 u (b) = b2 ,

(5.55) (5.56)

where αij and βij are real constants. We assume that the row vectors (α11 , α12 , β11 , β12 )

and

(α21 , α22 , β21 , β22 )

are linearly independent so that (5.55) and (5.56) represent two essentially different boundary conditions. There are two common kinds of boundary data. The first type is called the separated boundary conditions: β11 = β12 = α21 = α22 = 0, that is, B1 (u) = α11 u(a) + α12 u (a) = b1 ,

B2 (u) = β21 u(a) + β22 u (a) = b2 .

(5.57) (5.58)

240

Chapter 5

Integral and Differential Equations

Another kind of boundary condition, which often arises in practice, is called the periodic boundary conditions: u(a) = u(b),



u (a) = u (b).

(5.59) (5.60)

Example 5.8.1. Consider the operator L = D2 + ω2 with the separated bound-

ary conditions u(0) = u(1) = 0, 0 ≤ x ≤ 1.




Example 5.8.2. Consider the deflection of a string stretched under a constant tension T0 with fixed ends at x = a and x = b, and subject to a force distribution f (x). The differential equation and the boundary conditions for the vertical displacement u(x) of the string of density ρ(x) are d2 u = f (x), a ≤ x ≤ b; dx2 u(a) = u(b) = 0, where f (x) = gρ(x)/T0 and g is the acceleration due to gravity. In the context of boundary value problems when we deal with a differential operator we actually have in mind the whole system: the differential operator itself and the given boundary conditions. This convention is evident in the following definition.


Definition 5.8.3.

(Domain of a differential operator) Let L be a differential operator in L2 ([a, b]). The domain of L, denoted by D(L), is the subset of L2 ([a, b]) consisting of all functions for which the highest derivative in L is square integrable and satisfying the homogeneous boundary conditions. For example, if L is a second-order differential operator, then D(L) consists of all functions in L2 ([a, b]) which have square integrable second derivatives and satisfy the boundary conditions B1 (u) = B2 (u) = 0. Since differential operators are unbounded, adjoint operators are defined as in Section 4.11.

Definition 5.8.4.

(Adjoint of a differential operator) Let L be a differential operator in L2 ([a, b]). An operator L∗ is called the adjoint of L if $ % Lu, v = u, L∗ v (5.61) for all u ∈ D(L) and v ∈ D(L∗ ).

241

5.8 Ordinary Differential Equations

The adjoint of the differential operator in (5.54) can be found as follows:  b Lu, v = (Lu)(x)v(x)dx 

a

 a2 (x)u

(x) + a1 (x)u (x) + a0 (x)u(x) v(x)dx

b

= 

a

    b = a2 (x) u (x)v(x) − u(x)v (x) + a1 (x) − a 2 (x) u(x)v(x) a  b  

   + u(x) a2 (x)v(x) − a1 (x)v(x) + a0 (x)v(x) dx. a

If the right-hand side of this result is equal to u, L∗ v, we have  b %  

   $ ∗ u(x) a2 (x)v(x) − a1 (x)v(x) + a0 (x)v(x) dx u, L v = a

     b + a2 (x) u (x)v(x) − u(x)v (x) + a1 (x) − a 2 (x) u(x)v(x) a .

Thus, L∗ consists of the second-order differential operator   d2  d a2 (x)v(x) − a1 (x)v(x) + a0 (x)v(x) 2 dx dx

and some boundary terms. Therefore, the adjoint of L = a2

d2 d + a1 + a0 2 dx dx

is L∗ = a2

 

d  d2

+ 2a − a − a + a + a . 1 0 2 2 1 dx2 dx

(5.62)

The expression     J(u, v) = a2 vu − uv + a1 − a 2 uv

(5.63)

is called the bilinear concomitant of u and v . We can say that v ∈ D(L∗ ) if it is twice differentiable and [J(u, v)]ba = 0 for every u ∈ D(L). From the definition of J , it follows that [J(u, v)]ba = 0 for all u ∈ D(L) only if v satisfies two homogeneous boundary conditions B∗1 (v) = B∗2 (v) = 0. These are referred to as adjoint boundary conditions. In other words, corresponding to the homogeneous boundary value problem Lu = f ,

B1 (u) = B2 (u) = 0,

there is an adjoint boundary value problem L∗ v = f ,

B∗1 (v) = B∗2 (v) = 0.

242

Chapter 5

Integral and Differential Equations

Definition 5.8.5.

(Self-adjoint and formally self-adjoint differential operator) A differential operator L is called self-adjoint if L = L∗ and D(L) = D(L∗ ). A differential operator L is called formally self-adjoint if L = L∗ . It follows from (5.62) that if L is self-adjoint, then a 2 = a1 . Consequently, a selfadjoint differential operator L can be written as   d d a2 + a0 , L= (5.64) dx dx

and J(u, v) is then given by   J(u, v) = a2 u v − uv .

Example 5.8.6. We will consider the differential operator defined by Lu =

d2 u dx2

and the following boundary value problem: Lu = 0,

a ≤ x ≤ b,

B1 (u) = u(a) = 0

and

B2 (u) = u (a) = 0.

Clearly, L = L∗ . To determine the adjoint boundary conditions, we compute b  b  J(u, v) a = u (x)v(x) − u(x)v (x) a = u (b)v(b) − u(b)v (b) − v(a)u (a) + u(a)v (b) = u (b)v(b) − u(b)v (b). Hence [J(u, v)]ba = 0 for all u ∈ D (L) if B∗1 (v) = v(b) = 0 and

B∗2 (v) = −v (b) = 0.

Although this operator is formally self-adjoint, it is not self-adjoint because D(L) = D(L∗ ).


Example 5.8.7. Consider the boundary value problem d2 d u(x) + a1 (x) u(x) + a0 (x)u(x), dx2 dx u (1) = 0,

(Lu)(x) = a2 (x) u(0) = 0,

243

5.8 Ordinary Differential Equations

where a2 (x) = 0 on [0, 1]. Then L∗ = a2

  d  d2 + a

2 − a 1 + a0 . + 2a 2 − a1 2 dx dx

To find the adjoint boundary conditions, we compute    1  1  J(u, v) 0 = a2 (x) u (x)v(x) − u(x)v (x) + u(x)v(x) a1 (x) − a 2 (x) 0   = a1 (1) − a 2 (1) u(1)v(1) − a2 (1)v (1)u(1) − a2 (0)v(0)u (0). Hence, [J(u, v)]10 = 0 for all u ∈ D (L), if B∗1 = v(0) = 0

and

B∗2 (v) = v (1) −

(a1 (1) − a 2 (1))v(1) = 0. a2 (1)

Thus, in general, D (L) = D (L∗ ). However, if a 2 = a1 , then L = L∗ and B∗1 (v) = v(0) = 0 and

B∗2 (v) = v (1) = 0,

so that D (L) = D (L∗ ). That means that the boundary value problem is selfadjoint.
If an operator L of the form (Lu)(x) = a2 (x)

d2 d u(x) + a1 (x) u(x) + a0 (x)u(x), 2 dx dx

where a2 (x) = 0 on [a, b], is not self-adjoint, then it can be transformed into a self-adjoint form. We multiply L by , a1 (x) 1 exp dx a2 (x) a2 (x) to obtain , a1 (x) 1 exp dx (Lu)(x) a2 (x) a2 (x)  , , d a1 (x) a0 (x) a1 (x) d = exp dx u(x) + exp dx u(x), dx a2 (x) dx a2 (x) a2 (x) which is a formally self-adjoint operator. To study the second-order differential equations, it is convenient to express them in the form (Lu)(x) + λω(x)u(x) = 0,

a ≤ x ≤ b,

(5.65)

244

Chapter 5

Integral and Differential Equations

where L is self-adjoint, that is,   d d L= p + q, dx dx λ is a constant parameter, and ω is a given function, called the weight function. We assume ω(x) > 0 except possibly at isolated points where ω(x) = 0. For a given value of the parameter λ, a function uλ satisfying (5.65) and prescribed boundary conditions is called an eigenfunction corresponding to λ, which is then called an eigenvalue. Clearly, there is no guarantee that for every value of λ there exists an eigenfunction uλ .

Example 5.8.8. (Legendre operator and Legendre polynomials) The Legendre differential equation is ,  d  2 du 1−x + λu = 0, −1 ≤ x ≤ 1 dx dx (p(x) = 1 − x2 , q(x) = 0, ω(x) = 1). The eigenfunctions of this equation are called the Legendre polynomials: Pn (x) =

n 1 dn  2 x −1 , 2n n! dxn

and the corresponding eigenvalues are λn = n(n + 1), n = 1, 2, 3, . . . . By the Legendre operator, we mean ,  d  2 du 1−x . Lu = −
dx dx

Example 5.8.9. (Associated Legendre operator and associated Legendre functions) The eigenfunctions of the associated Legendre differential equation ,  d  m2 u 2 du 1−x − + λu = 0, −1 ≤ x ≤ 1 dx dx 1 − x2 (p(x) = 1 − x2 , q(x) = −m2 /(1 − x2 ), ω(x) = 1) are the associated Legendre functions:  m/2 dm Pn (x), n > m, Pnm (x) = 1 − x2 dxm and the corresponding eigenvalues are λn = n(n + 1), n = 1, 2, 3, . . . . Note that the operator ,  du m2 u d  1 − x2 + Lu = − dx dx 1 − x2 is positive definite.




5.8 Ordinary Differential Equations

245

Example 5.8.10. (Chebyshev operator and Chebyshev polynomials) The Chebyshev differential equation is , d du 2 2 1−x 1−x + λu = 0, −1 ≤ x ≤ 1 dx dx √ (p(x) = 1 − x2 , q(x) = 0, ω(x) = (1 − x2 )−1/2 ). The eigenfunctions   Tn (x) = cos n cos−1 x are called the Chebyshev polynomials. The corresponding eigenvalues are λn = n2 . The Chebyshev operator is defined by , du d 2 . 1−x Lu = −
dx dx

Example 5.8.11. (Jacobi operator and Jacobi polynomials) The differential equation (1 − x)

−α

(1 + x)

−β

, d α+1 β+1 du (1 − x) (1 + x) + λu = 0, dx dx

−1 ≤ x ≤ 1,

with α > −1 and β > −1, is called the Jacobi differential equation (p(x) = (1 − x)α+1 (1 + x)β+1 , q(x) = 0, ω(x) = (1 − x)α (1 + x)β ). The functions Pn(α,β) (x) =

 (−1)n  (1 − x)α+n (1 + x)β+n , n 2 n!

called the Jacobi polynomials, are the eigenfunctions of the Jacobi equation. The eigenvalues are λn = n(n + α + β + 1). The Jacobi operator is defined by , d α+1 β+1 du Lu = − (1 − x) (1 + x) . dx dx If α = β = 0, the Jacobi equation becomes the Legendre equation. For α = β = − 12 , it reduces to the Chebyshev equation. If α = β = γ − 12 , then we get γ the Gegenbauer polynomials Cn (x).


Example 5.8.12. (Laguerre operator and Laguerre polynomials) The differential equation

, d −x du e xe + λu = 0, dx dx x

0 < x < ∞,

is called the Laguerre differential equation. The Laguerre operator is , d −x du xe . Lu = − dx dx

246

Chapter 5

Integral and Differential Equations

Its eigenvalues are λn = n and the corresponding eigenfunctions, called the Laguerre polynomials, are defined by n   1 x dn  n −x   n xk Ln (x) = e n x e = . k k! n! dx




k=1

Example 5.8.13. (Associated Laguerre operator and associated Laguerre functions) The equation , −α x d α+1 −x du x e x e + λu = 0, 0 ≤ x ≤ ∞, dx dx is called the associated Laguerre equation. The associated Laguerre functions Lαn (x) = (−1)n

 ex x−α dn  −x n+α  dα  (x) = L e x n+α dxα n! dxn

are the eigenfunctions of the differential operator , d α+1 −x du Lu = − x e dx dx with the eigenvalues λn = n − α.




Example 5.8.14. (Bessel operator and Bessel functions) By the Bessel equation we mean

, ν2 1 d du x − 2 u + λu = 0, x dx dx x

0 < x < a.

The corresponding operator is Lu = −

, ν2 d du x + u, dx dx x

and its eigenfunctions are the Bessel functions Jν (kn x) with the eigenvalues λn = kn2 , n = 1, 2, 3, . . . , where akn are positive zeros of Jν (t). The Bessel operator is positive definite, and its eigenfunctions form a complete orthogonal system in L2 [(0, a]) with respect to the inner product with the weight function ω(x) = x.


Example 5.8.15. (Hermite operator and Hermite polynomials) The differential equation , d −x2 du e + λu = 0, e dx dx x2

−∞ < x < ∞,

5.9 Sturm–Liouville Systems

and the operator

247

, d −x2 du e Lu = − dx dx

are called the Hermite equation and Hermite operator, respectively. The eigenfunctions n 2 d  2 Hn (x) = (−1)n ex n e−x dx are called the Hermite polynomials; the eigenvalues are λn = 2n.


5.9 Sturm–Liouville Systems In this section we are going to concentrate on a special class of boundary value problems referred to as the Sturm–Liouville system (Jacques Charles François Sturm (1803–1855), Joseph Liouville (1809–1882)).

Definition 5.9.1. (Regular Sturm–Liouville system) A regular Sturm–Liouville system consists of an ordinary differential equation ,   du d p(x) + q(x) + λω(x) u = 0, a ≤ x ≤ b, (5.66) dx dx with the boundary conditions a1 u(a) + a2 u (a) = 0,

(5.67a)

b1 u(b) + b2 u (b) = 0,

(5.67b)

where p, q, and ω are continuous real-valued functions on [a, b], p and ω are positive in [a, b], p (x) exists and is continuous in [a, b], a1 , a2 , b1 , b2 are given real numbers such that a21 + a22 > 0 and b21 + b22 > 0.

Definition 5.9.2. (Singular Sturm–Liouville system) Suppose p, q, and ω are functions defined in [a, b] satisfying all the conditions in Definition 5.9.1 except that p is only assumed to be positive in (a, b), and vanishes at one or both end points of the interval [a, b]. By the singular Sturm–Liouville system, we mean the system consisting of the differential equation (5.66) in the open or semiopen interval with boundary conditions (a) u is bounded on (a, b),

248

Chapter 5

Integral and Differential Equations

(b) if p does not vanish at an end point, then p satisfies boundary conditions of the form (5.67a) and (5.67b) at that end point.

Example 5.9.3. Consider the regular Sturm–Liouville system u

+ λu = 0,

0 ≤ x ≤ π,

u(0) = u(π ) = 0. √ Suppose λ < 0 and let ν = |λ|. Then, the general solution of the equation is u(x) = Aeνx + Be−νx . This solution satisfies the given boundary conditions if and only if A + B = 0, Ae

νπ

+ Be−νπ = 0.

Since ν > 0, eνπ = e−νπ and therefore the only solution is A = B = 0. This means that the system has no nonzero solutions if λ < 0. In other words, there are no negative eigenvalues. A similar argument shows that λ = 0 is not an eigenvalue. However, when λ > 0, then the solutions of the equation are √ √ u(x) = A cos λx + B sin λx. The boundary conditions give A=0

and

√ B sin λπ = 0.

Since λ = 0 and B = 0 yields trivial solution, we must have B = 0 and √ sin λπ = 0. Hence, the eigenvalues are λn = n2 , n = 1, 2, . . . , and the eigenfunctions are un (x) = sin nx. Note that λn → ∞ as n → ∞, unlike the case of self-adjoint compact operators when the eigenvalues converge to 0; see Theorem 4.9.19. Section 5.10, particularly Theorem 5.10.4, will explain this.
Another type of problem that often occurs in practice is the periodic Sturm– Liouville system: ,   du d p(x) + q(x) + λω(x) u = 0, a ≤ x ≤ b, dx dx

5.9 Sturm–Liouville Systems

p(a) = p(b),

u(a) = u(b),

249

u (a) = u (b).

Example 5.9.4. Find the eigenvalues and eigenfunctions of the following periodic Sturm–Liouville system u

+ λu = 0, u(−π ) = u(π ),

−π ≤ x ≤ π, u (−π ) = u (π ).

Note that here p(x) = 1, and hence p(−π ) = p(π ). When λ > 0 the general solution of the equation is √ √ u(x) = A cos λx + B sin λx. Using the boundary conditions, we get √ 2B sin λπ = 0, √ √ 2A λ sin λπ = 0. Thus, for nontrivial solutions, we must have √ sin λπ = 0. The equation is satisfied if λ = λn = n2 ,

n = 1, 2, 3, . . . .

For every eigenvalue λn = n2 , we have two linearly independent solutions cos nx and sin nx. It can be readily shown that the system has no negative eigenvalues. However, λ = 0 is an eigenvalue and the corresponding eigenfunction is the constant function u(x) = 1. Thus, the eigenvalues are 0, 1, 4, . . . , n2 , . . . , and the corresponding eigenfunctions are 1, cos x, sin x, cos 2x, sin 2x, . . . , cos nx, sin nx, . . . . Throughout the remainder of this section, L denotes the differential operator in the Sturm–Liouville differential equation, that is, , du d p(x) + q(x)u. Lu = dx dx For the regular Sturm–Liouville system, we denote by D (L) the domain of L, that is, D (L) is the space of all complex-valued functions u defined on [a, b]

250

Chapter 5

Integral and Differential Equations

for which u

belongs to L2 ([a, b]) and that satisfy boundary conditions (5.67a) and (5.67b). Then, we have   L : D (L) → L2 [a, b] . For the singular Sturm–Liouville system, we need only to replace (5.67a) and (5.67b) by (a) u is bounded on (a, b), (b) b1 u(b) + b2 u (b) = 0, where b1 and b2 are real constants such that b21 + b22 > 0.




The following identity is attributed to Joseph-Louis Lagrange (1736–1813).

Theorem 5.9.5. (Lagrange’s identity) For any u, v ∈ D(L), we have ,  dv du d p u −v . uLv − vLu = dx dx dx

(5.68)

Proof: We have , , d du d dv p + quv − v p − quv dx dx dx dx ,  d dv du = p u −v . dx dx dx

uLv − vLu = u




Theorem 5.9.6. (Abel’s formula) If u and v are two solutions of the equation Lu + λωu = 0

(5.69)

in [a, b], then p(x)W(x; u, v) = constant, where W is the Wronskian defined by   u(x) W(x; u, v) =  v(x)

 u (x)  . v (x) 

The Wronskian is named after Josef Hoëné de Wronski (1778–1853). Proof: Since u and v are solutions of (5.69), we have ,   d du p(x) + q(x) + λω(x) u = 0, dx dx ,   dv d p(x) + q(x) + λω(x) v = 0. dx dx

5.9 Sturm–Liouville Systems

251

Multiplying the first equation by v and the second by u, and then subtracting, we obtain , , d du d dv p −v p = 0. u dx dx dx dx By integrating this equation from a to x, we find     p(x) u(x)v (x) − u (x)v(x) − p(a) u(a)v (a) − u (a)v(a) = constant. This is Abel’s formula.




Theorem 5.9.7. Eigenfunctions of a regular Sturm–Liouville system are unique except for a constant factor.

Proof: Suppose u and v are eigenfunctions corresponding to the same eigenvalue λ. According to Abel’s formula, we have p(x)W(x; u, v) = constant. Since p > 0, if W(x; u, v) vanishes at a point in [a, b], then it vanishes everywhere in [a, b]. From the boundary conditions, we have a1 u(a) + a2 u (a) = 0, a1 v(a) + a2 v (a) = 0. Since a1 and a2 are not both zero, we get    u(a) u (a)    = 0. W(a; u, v) =  v(a) v (a)  Therefore, W(x; u, v) = 0 for all x ∈ [a, b], which proves linear dependence of u and v.


Theorem 5.9.8. For any u, v ∈ D(L), we have Lu, v = u, Lv, where  ,  denotes the inner product of L2 ([a, b]). In other words, L is a selfadjoint operator.

Proof: Since all constants involved in the boundary conditions of a Sturm–

Liouville system are real, if v ∈ D (L), then v ∈ D (L). Also, since p, q, and ω are real-valued, Lv = Lv. Consequently, 

b

Lu, v − u, Lv = a

 b (vLu − uLv)dx = p(u v − uv ) a ,

(5.70)

252

Chapter 5

Integral and Differential Equations

by Lagrange’s identity (5.68). We will show that the last term in the preceding equality vanishes for both the regular and singular systems. If p(a) = 0, the result follows immediately. If p(a) > 0, then u and v satisfy boundary conditions of the form (5.67a) and (5.67b) at x = a. That is, , -, u(a) u (a) a1 = 0.

v(a) v (a) a2 Since a1 and a2 are not both zero, we have u(a)v (a) − v(a)u (a) = 0. A similar argument can be applied to the other end point x = b, so that we conclude b  p(uv − vu ) a = 0.


Theorem 5.9.9. Eigenvalues of a Sturm–Liouville system are real. Proof: Let λ be an eigenvalue of a Sturm–Liouville system, and let u be the corresponding eigenfunction. This means that u = 0 and Lu = −λωu. Then 0 = Lu, u − u, Lu = −λωu, u − u, −λωu   2   b = λ−λ ω(x)u(x) dx. a

Since ω(x) > 0 in [a, b] and u = 0, the integral is a positive number. Therefore λ = λ, completing the proof.
This theorem states that all eigenvalues of a regular Sturm–Liouville system are real, but it does not guarantee that an eigenvalue exists. It is proved in Section 5.10 that a regular Sturm–Liouville system has an infinite sequence of eigenvalues.

Theorem 5.9.10. Eigenfunctions corresponding to distinct eigenvalues of a Sturm–Liouville system are orthogonal with respect to the inner product with the weight function ω(x).

Proof: Suppose u1 and u2 are eigenfunctions corresponding to eigenvalues λ1 and λ2 , λ1 = λ2 . Thus, Lu1 = −λωu1

and

Lu2 = −λ2 ωu2 .

Hence, u1 Lu2 − u2 Lu1 = −λ2 ωu1 u2 + λ1 ωu1 u2 = (λ1 − λ2 )ωu1 u2 .

(5.71)

5.10 Inverse Differential Operators

By Theorem 5.9.5, we have u1 Lu2 − u2 Lu1 =

,  d du2 du1 p u1 − u2 . dx dx dx

253

(5.72)

Combining (5.71) and (5.72) and integrating from a to b, we get  b ω(x)u1 (x)u2 (x)dx (λ1 − λ2 ) ,

a

 du2 (x) du1 (x) b = p(x) u1 (x) = 0, − u2 (x) dx dx a by boundary conditions (5.67a) and (5.67b). Since λ1 = λ2 , we conclude  b ω(x)u1 (x)u2 (x)dx = 0.
a

5.10 Inverse Differential Operators and Green’s Functions A typical boundary value problem for an ordinary differential equation can be written in the operator form as Lu = f .

(5.73)

We seek a solution u, which satisfies this equation and the given boundary conditions. If D (L) is defined as the space of functions satisfying those boundary conditions, then the problem reduces to finding a solution of (5.73) in D (L). One way to approach the problem is by looking for the inverse operator L−1 . If it is possible to find L−1 , then the solution of (5.73) can be obtained as u = L−1 (f ). It turns out that it is possible in many important cases, and the inverse operator is an integral operator of the form  b  −1  G(x, t)f (t)dt. u = L f (x) = a

The function G is called the Green’s function (George Green (1793–1841)) of the operator L. Existence of the Green’s function and its determination is not a simple problem. However, we will examine the question more closely in the case of Sturm–Liouville systems.

Theorem 5.10.1. Suppose λ = 0 is not an eigenvalue of the following regular Sturm–Liouville system: , du d p(x) + q(x)u = f (x), a ≤ x ≤ b, (5.74) Lu = dx dx

254

Chapter 5

Integral and Differential Equations

with the homogeneous boundary conditions a1 u(a) + a2 u (a) = 0,

(5.75)

b1 u(b) + b2 u (b) = 0,

(5.76)

where p and q are continuous real-valued functions on [a, b], p is positive in [a, b], p (x) exists and is continuous in [a, b], a1 , a2 , b1 , b2 are given real numbers such that a21 + a22 > 0 and b21 + b22 > 0. Then, for any f ∈ C ([a, b]), the system has a unique solution 

b

u(x) =

G(x, t)f (t)dt, a

where G(x, t) is the Green’s function given by  u2 (x)u1 (t)    p(t)W(t) for a ≤ t < x, G(x, t) =  u (x)u2 (t)   1 for x < t ≤ b, p(t)W(t) where u1 and u2 are nonzero solutions of the homogeneous system , d du p(x) + q(x)u = 0 dx dx with boundary conditions (5.75) and (5.76), respectively, and W(t) = u1 (t)u 2 (t) − u2 (t)u 1 (t) is the Wronskian.

Proof: According to the theory of ordinary differential equations, the general solution of (5.74) is of the form u(x) = c1 u1 (x) + c2 u2 (x) + up (x),

(5.77)

where c1 and c2 are constants, u1 and u2 are two linearly independent solutions of the homogeneous equation Lu = 0, and up is any particular solution of (5.74). The particular solution up can be found by the method of variation of parameters. Thus, we are looking for a solution in the form up (x) = v1 (x)u1 (x) + v2 (x)u2 (x),

(5.78)

where v1 and v2 are functions to be determined. Since there are infinitely many pairs of functions v1 and v2 for which up satisfies (5.74), we impose a

5.10 Inverse Differential Operators

255

second condition: v1 u1 + v2 u2 = 0.

(5.79)

We now have u p = v1 u 1 + v2 u 2 and u

p = v1 u

1 + v2 u

2 + v1 u 1 + v2 u 2 . Substituting into (5.74), we get       v1 pu

1 + p u 1 + qu1 + v2 pu

2 + p u 2 + qu2 + p v1 u 1 + v2 u 2 = f . Since u1 and u2 are solutions of the homogeneous equation, the first two terms vanish so that the above result becomes f v1 u 1 + v2 u 2 = . p

(5.80)

Solving (5.79) and (5.80) for v1 and v2 , we obtain v1 (x) = −

f (x)u2 (x) p(x)W(x; u1 , u2 )

and

v2 (x) =

f (x)u1 (x) . p(x)W(x; u1 , u2 )

(5.81)

We will show that the Wronskian does not vanish at any point of [a, b]. Indeed, suppose that    u1 u2 

  W = u1 u2 − u2 u1 =  u1 u 2  vanishes at some ξ ∈ [a, b]. Then the system of equations αu1 (ξ ) + βu2 (ξ ) = 0, αu 1 (ξ ) + βu 2 (ξ ) = 0 has a nontrivial solution such that α and β are not both zero. Then the function g = αu1 + βu2 is a solution of the initial value problem , d du p(x) + q(x)u = 0, g(ξ ) = g (ξ ) = 0. dx dx But we know that the above problem has only the trivial solution, and thus g = 0. This means that u1 and u2 are linearly dependent, contrary to the assumption. By Abel’s formula (Theorem 5.9.6), p(x)W(x; u1 , u2 ) is a constant. Since W does not vanish in [a, b] and p is assumed to be positive, the constant is

256

Chapter 5

Integral and Differential Equations

not zero. Denote c=

1 . p(x)W(x; u1 , u2 )

Now, by integrating differential equations (5.81), we get   v1 (x) = − cf (x)u2 (x)dx and v2 (x) = cf (x)u1 (x)dx, and finally,  up (x) = −cu1 (x)

x

 f (t)u2 (t)dt + cu2 (x)

b



x

=

f (t)u1 (t)dt

(5.82)

a



b

cu2 (x)u1 (t)f (t)dt +

a

x

cu1 (x)u2 (t)f (t)dt.

(5.83)

x

Consequently, if we define the Green’s function as  cu2 (x)u1 (t) for a ≤ t < x, G(x, t) = cu1 (x)u2 (t) for x < t ≤ b, we can write



(5.84)

b

up (x) =

G(x, t)f (t)dt a

provided the integral exists. This follows immediately from the continuity of G. The proof of continuity of G is left as an exercise.
Denote by T the integral operator defined in Theorem 5.10.1, that is, 

b

(Tf )(x) =

G(x, t)f (t)dt.

(5.85)

a

We are going to examine properties of T.

Theorem 5.10.2. The operator T defined by (5.85) is a self-adjoint compact op-

erator from L2 ([a, b]) into C ([a, b]).

Proof: Compactness of integral operators was discussed in Example 4.8.4. In Example 4.4.6 we prove that an integral operator of the form (5.85) is self-adjoint if G(x, t) = G(t, x). It is easy to see that, in this case, the condition is satisfied. Finally, continuity of Tf follows from the continuity of G.
Operator T, as a compact, self-adjoint operator, admits spectral representation (see Corollary 4.10.2). The following two theorems describe the connection between eigenvalues and eigenfunctions of the regular Sturm–Liouville operator L and the corresponding integral operator T.

5.10 Inverse Differential Operators

257

Theorem 5.10.3. If λ = 0 is not an eigenvalue of the Sturm–Liouville system defined in Theorem 5.10.1, then λ = 0 is not an eigenvalue of the integral operator T defined by (5.85).

Proof: Suppose Tf = 0. Then ,  x  x d −cu1 (x) 0 = (Tf ) (x) = f (t)u2 (t)dt + cu2 (x) f (t)u1 (t)dt dx b a   x

= c −u1 (x) f (t)u2 (t)dt + u1 (x)u2 (x)f (x)

+ u 2 (x)  =c

b



x



f (t)u1 (t)dt − u1 (x)u2 (x)f (x)

a

−u 1 (x)



x

f (t)u2 (t)dt b

+ u 2 (x)



x

 f (t)u1 (t)dt .

a

Therefore, we have the following system of equations:  x  x f (t)u2 (t)dt − u2 (x) f (t)u1 (t)dt = 0, u1 (x) b

u 1 (x)



a

x

b

Since the determinant

f (t)u2 (t)dt − u 2 (x)   u1  u 1



x

f (t)u1 (t)dt = 0.

a

 u2  u 2 

does not vanish at any point of [a, b] (see the proof of Theorem 5.10.1), we conclude  x  x f (t)u2 (t)dt = 0 and f (t)u1 (t)dt = 0 b

a

for all x ∈ [a, b]. This implies f = 0, and thus, the equation Tf = 0 has only the trivial solution.


Theorem 5.10.4. Under the assumptions of Theorem 5.10.1, λ is an eigenvalue of L if and only if 1/λ is an eigenvalue of T. Moreover, if f is an eigenfunction of L corresponding to the eigenvalue λ, then f is an eigenfunction of T corresponding to the eigenvalue 1/λ.

Proof: Suppose Lf = λf for some nonzero f in the domain of L. By Theorem 5.10.1, we have f = T(λf )

258

Chapter 5

Integral and Differential Equations

or, equivalently, since λ = 0, 1 Tf = f . λ This shows that 1/λ is an eigenvalue of T and f is the corresponding eigenfunction. Conversely, if f is an eigenfunction of T corresponding to λ, f = 0 and λ = 0, then Tf = λf , and hence, f = L(Tf ) = L(λf ) = λLf . Therefore, 1/λ is an eigenvalue of L and the corresponding eigenfunction is f .


5.11 The Fourier Transform In this section, we introduce the Fourier transform in L2 (R) and discuss its basic properties. The definition of the transform in L2 (R) is not trivial. The integral  ∞ 1 e−iωx f (x)dx √ 2π −∞ cannot be used as a definition of the Fourier transform in L2 (R) because not all functions in L2 (R) are integrable. It is, however, possible to extend the Fourier transform from L1 (R) ∩ L2 (R) onto L2 (R). In the first part of this section, we discuss properties of the Fourier transform in L1 (R). Then, we show that the extension onto L2 (R) is possible and study properties of that extension. Let f be an integrable function on R. Consider the integral  ∞ e−iωx f (x)dx, ω ∈ R. (5.86) −∞

Since the function g(x) = e−iωx is continuous, the product e−iωx f (x) is a locally integrable function for every ω ∈ R (see Theorem 2.9.4). Moreover, since |e−iωx | = 1 for all ω, x ∈ R, we have |e−iωx f (x)| = |f (x)| and thus, by Theorem 2.9.5, the integral (5.86) exists for all ω ∈ R.

(Fourier transform in L1 (R)) Let f ∈ L (R). The function fˆ defined by  ∞ ˆf (ω) = √1 e−iωx f (x)dx 2π −∞

Definition 5.11.1. 1

(5.87)

5.11 The Fourier Transform

259

is called the Fourier transform of f . In some books the Fourier transform is defined without the factor √12π . Anvariation is the definition without the “−” sign in the exponent, that is, other ∞ iωx e f (x)dx. These details do not change the theory of the Fourier trans−∞ forms at all. Instead of “fˆ ” the notation “F {f (x)}” is also used. The latter is especially convenient if instead of a letter “f ” or “g” we want to use a formula describing 2 a function, for example, F (e−x }. We will use both symbols freely.

Example 5.11.2. (a) Let α > 0. Then
 1 F e−α|x| = √ 2π



−∞

e−iωx e−α|x| dx

 ∞ 1 e dx + √ e−x(α+iω) dx 2π 0 −∞ - . , 1 α 1 2 1 + = =√ . 2 + ω2 α − iω α + iω π α 2π 1 =√ 2π



∞

0

x(α−iω)

(b) ,   iω 2 ω2 exp −α x + − dx 2α 4α −∞  ∞ 1 1 2 2 2 = √ e−ω /(4α) e−αu du = √ e−ω /(4α) . 2π 2α −∞


1 2 F e−αx = √ 2π



∞




The following theorem is an immediate consequence of Definition 5.11.1.

Theorem 5.11.3. (Linearity) Let f , g ∈ L1 (R) and α, β ∈ C. Then F (αf + βg) = α F (f ) + β F (g).

Theorem 5.11.4. The Fourier transform of an integrable function is a continuous function.

Proof: Let f ∈ L1 (R). For any ω, λ ∈ R, we have

   ∞       −iωx −iλx ˆf (ω + λ) − fˆ (ω) =  √1 e − 1 f (x)dx e  2π −∞  ∞  −iλx   1 e ≤√ − 1f (x)dx. 2π −∞

(5.88)

260

Chapter 5

Integral and Differential Equations

Since

 −iλx     e − 1f (x) ≤ 2f (x)

and

  lim e−iλx − 1 = 0 for all x ∈ R,

λ→0

we conclude 1 lim √ λ→0 2π



∞ −∞

  e−iλx − 1f (x)dx = 0

by the Dominated Convergence Theorem. This proves continuity of fˆ . (Actually, since the inequality in (5.88) is independent of ω, we have proved that fˆ is uniformly continuous.)
∞ Recall from Section 2.6 that the integral −∞ |f (x)|dx defines a norm in L1 (R). This norm will be denoted by · 1 , that is,  f 1 =

∞ −∞

 f (x)dx

for f ∈ L1 (R).

Theorem 5.11.5. If f1 , f2 , . . . ∈ L1 (R) and fn − f 1 → 0 as n → ∞, then fˆn →

fˆ uniformly on R.

Proof: First notice that   ˆf (ω) ≤ √1 2π



∞

 e−iωx f (x)dx ≤ √1 2π −∞



∞ −∞

 f (x)dx,

for every ω ∈ R. Thus,   1 supˆfn (ω) − fˆ (ω) ≤ √ 2π ω∈R



∞ −∞

 fn (x) − f (x)dx = fn − f 1 → 0,

proving the theorem.




Theorem 5.11.6. (Riemann–Lebesgue lemma) If f ∈ L1 (R), then   lim ˆf (ω) = 0.

|ω|→∞

Proof: Since e−iωx = −e−iωx−iπ , we have 1 fˆ (ω) = − √ 2π



∞

−∞

e

−iω(x+π/ω)

1 f (x)dx = − √ 2π



∞ −∞

e

−iωx

  π f x− dx. ω

5.11 The Fourier Transform

261

Hence, ,    ∞  ∞ 1 π −iωx −iωx ˆf (ω) = 1 √1 dx e f (x)dx − √ e f x− 2 ω 2π −∞ 2π −∞     ∞ 1 1 π −iωx = √ dx. f (x) − f x − e 2 2π −∞ ω We thus have   ˆf (ω) ≤ √1 2 2π Since

 lim

|ω|→∞



    f (x) − f x − π dx.  ω  −∞ ∞

    f (x) − f x − π dx = 0,  ω  −∞ ∞

by Theorem 2.4.3, we have lim|ω|→∞ |fˆ (ω)| = 0.




Note that the space C0 (R) of all continuous functions on R, which vanish at infinity (that is, such that lim|x|→∞ f (x) = 0), is a normed space with respect to the norm defined by   f = supf (x). x∈R

Theorems 5.11.3 to 5.11.6 show that the Fourier transform is a continuous linear operator from L1 (R) into C0 (R).

Theorem 5.11.7. Let f ∈ L1 (R). Then 
(a) F eiαx f (x) = fˆ (ω − α) (translation), 
(b) F f (x − x0 ) = fˆ (ω)e−iωx0 (shifting),  
 1 ω ˆ (c) F f (αx) = f , α > 0 (scaling), α α 

 (d) F f (x) = F f (−x) (conjugate). Proof follows easily from Definition 5.11.1.

Example 5.11.8. (Modulated Gaussian function) If 1 2

f (x) = eiω0 x− 2 x , then 1 2 fˆ (ω) = e− 2 (ω−ω0 ) .

262

Chapter 5

Integral and Differential Equations

Figure 5.2 Modulated Gaussian and its Fourier transform (ω0 = 4). This readily follows from Example 5.11.2(b) combined with the translation property of the Fourier transform. The graphs of Re f (x) and fˆ (ω) are drawn in Figure 5.2. The Gaussian function is named after Johann Carl Friedrich Gauss (1777–1855).


Theorem 5.11.9. If f is a continuous piecewise differentiable function, f , f ∈

L1 (R), and lim|x|→∞ f (x) = 0, then
 F f = iωF {f }.

Proof: Simple integration by parts gives 1 √ 2π



∞ −∞

f (x)e

−iωx

∞ 1  iω dx = √ f (x)e−iωx −∞ + √ 2π 2π = iωfˆ (ω).



∞ −∞

f (x)e−iωx dx


Corollary 5.11.10. If f is a continuous function, n-times piecewise differentiable,

and f , f , . . . , f (n) ∈ L1 (R), and

lim f (k) (x) = 0 for k = 0, . . . , n − 1,

|x|→∞

then


 F f (n) = (iω)n F {f }.

Because of our definition of the Fourier transform, it is better to redefine the convolution of two functions f , g ∈ L1 (R) as follows:  ∞ 1 f (x − u)g(u)du. (f ∗ g)(x) = √ 2π −∞

5.11 The Fourier Transform

263

The main reason is the simplicity of the formula in the next theorem.

Theorem 5.11.11. (Convolution theorem) Let f , g ∈ L1 (R). Then F (f ∗ g) = F (f )F (g).

Proof: Let f , g ∈ L1 (R) and h = f ∗ g. Then h ∈ L1 (R), by Theorem 2.15.1, and we have 1 ˆ h(ω) =√ 2π = = = =



∞

−∞  ∞

h(x)e−iωx dx

 ∞ 1 −iωx 1 e f (x − u)g(u)dudx √ √ 2π −∞ 2π −∞  ∞  ∞ 1 g(u) e−iωx f (x − u)dxdu 2π −∞ −∞  ∞  ∞ 1 g(u) e−iω(x+u) f (x)dxdu 2π −∞ −∞  ∞  ∞ 1 1 −iωu g(u)e du √ e−iωx f (x)dx √ 2π −∞ 2π −∞

= gˆ (ω)fˆ (ω).




We will now discuss the extension of the Fourier transform onto L2 (R). In the following theorem, and in the remaining part of this section, · 2 denotes the norm in L2 (R), that is, + ∞  f (x)2 dx for f ∈ L2 (R). f 2 = −∞

Theorem 5.11.12. Let f be a continuous function on R vanishing outside a bounded interval. Then fˆ ∈ L2 (R) and   ˆf  = f 2 . 2 Proof: Suppose first that f vanishes outside the interval [−π, π ]. Using Parseval’s formula for the orthonormal sequence of functions on [−π, π ], 1 ϕn (x) = √ e−inx , 2π

n = 0, ±1, ±2, . . . ,

we get f 22

 ∞    1  =  √2π n=−∞

∞

−∞

e

−inx

2 ∞     ˆf (n)2 . f (x)dx = n=−∞

264

Chapter 5

Integral and Differential Equations

Since this inequality holds also for g(x) = e−iξ x f (x) instead of f (x), we obtain ∞    2 ˆf (n + ξ )2 , f 2 = n=−∞

in view of to 1 yields

f 22

f 22

=

g 22 .

Integration of both sides with respect to ξ from 0

 ∞  1   2 ˆ  = f (n + ξ ) dξ = n=−∞ 0

∞ −∞

   ˆf (ξ )2 dξ = ˆf 2 . 2

If f does not vanish outside [−π, π ], then we take a positive number λ for which the function g(x) = f (λx) vanishes outside [−π, π ]. Then   1 x gˆ (x) = fˆ , λ λ and thus   2  ∞     1 ˆ ξ   ˆf (ξ )2 dx = ˆf 2 . =λ dξ = f  2 λ  −∞ λ −∞ 

f 22

= λ g 22

= λ ˆg 22

∞




The space of all continuous functions on R with compact support is dense in L2 (R). Theorem 5.11.12 shows that the Fourier transform is a continuous mapping from that space into L2 (R). Since the mapping is linear, it has a unique extension to a linear mapping from L2 (R) into itself. This extension will be called the Fourier transform on L2 (R).

(Fourier transform in L2 (R)) Let f ∈ L2 (R), and let (ϕn ) be a sequence of continuous functions with compact support convergent to f in L2 (R), that is, f − ϕn 2 → 0. The Fourier transform of f is defined by

Definition 5.11.13.

fˆ = lim ϕˆ n , n→∞

(5.89)

where the limit is with respect to the norm in L2 (R). Theorem 5.11.12 guarantees that the limit exists and is independent of a particular sequence approximating f . It is important to remember that the convergence in L2 (R) does not imply pointwise convergence, and therefore, the Fourier transform of a square integrable function is not defined at every point, unlike the Fourier transform of an integrable function. The Fourier transform

5.11 The Fourier Transform

265

of a square integrable function is defined almost everywhere. For this reason we cannot say that if f ∈ L1 (R) ∩ L2 (R), then the Fourier transform defined by (5.87) and the one defined by (5.89) are equal. To be precise, we should say that the function defined by (5.87) belongs to the equivalence class of square integrable functions defined by (5.89). In spite of this difference, we will use the same symbol to denote both transforms. It will not cause any misunderstanding. The following theorem is an immediate consequence of Definition 5.11.13 and Theorem 5.11.12.

Theorem 5.11.14. (Parseval’s relation) If f ∈ L2 (R), then   ˆf  = f 2 . 2 In physical problems, the quantity f 2 is a measure of energy, and fˆ 2 represents the power spectrum of f .

Theorem 5.11.15. Let f ∈ L2 (R). Then



1 fˆ (ω) = lim √ n→∞ 2π

n −n

e−iωx f (x)dx,

(5.90)

where the convergence is with respect to the norm in L2 (R).

Proof: For n = 1, 2, 3, . . . define



fn (x) =

if |x| < n, if |x| ≥ n.

f (x) 0

Then f − fn 2 → 0, and thus, fˆ − fˆn 2 → 0 as n → ∞.




Theorem 5.11.16. (Weak Parseval’s relation) If f , g ∈ L2 (R), then 

∞ −∞



f (x)ˆg (x)dx =

∞

−∞

fˆ (x)g(x)dx.

Proof: For n = 1, 2, 3, . . . define



fn (x) = and

 gn (x) =

f (x) 0

if |x| < n, if |x| ≥ n,

g(x) 0

if |x| < n, if |x| ≥ n.

Since 1 fˆm (x) = √ 2π



∞ −∞

e−ixξ fm (ξ )dξ,

(5.91)

266

Chapter 5

we have



Integral and Differential Equations

∞

1 fˆm (x)gn (x)dx = √ 2π −∞





∞ −∞

gn (x)

∞

−∞

e−ixξ fm (ξ )dξ dx.

The function e−ixξ gn (x)fm (ξ ) is integrable over R2 , and thus the Fubini theorem can be applied. Consequently,  ∞  ∞  ∞ 1 fˆm (x)gn (x)dx = √ fm (ξ ) e−ixξ gn (x)dxdξ 2π −∞ −∞ −∞  ∞ = fm (ξ )ˆgn (ξ )dξ. −∞

Since g − gn 2 → 0 and ˆg − gˆn 2 → 0 as n → ∞, we obtain  ∞  ∞ fm (x)ˆg (x)dx, fˆm (x)g(x)dx = −∞

−∞

by continuity of inner product. For the same reason, by letting m → ∞, we get  ∞  ∞ ˆf (x)g(x)dx = f (x)ˆg (x)dx.
−∞

−∞

The following technical lemma will be useful in the proof of the important inversion theorem for the Fourier transform in L2 (R).

Lemma 5.11.17. Let f ∈ L2 (R) and let g = fˆ . Then f = gˆ . Proof: From Theorems 5.11.14 and 5.11.16 and equality g = fˆ , we obtain $ % $ % $ %  2 f , gˆ = fˆ , g = fˆ , fˆ = ˆf 2 = f 22 .

(5.92)

$ % f , gˆ = f 22 .

(5.93)

 2 ˆg 22 = g 22 = ˆf 2 = f 22 .

(5.94)

Hence also

Finally, by Parseval’s equality,

Using (5.92) to (5.94), we get   $ % $ % $ % f − gˆ 2 = f − gˆ , f − gˆ = f 2 − f , gˆ − f , gˆ + ˆg 2 = 0. 2 2 2

5.11 The Fourier Transform

This shows that f = gˆ .

267


Theorem 5.11.18. (Inversion of the Fourier transform in L2 (R)) Let f ∈

L2 (R). Then

1 f (x) = lim √ n→∞ 2π



n −n

eiωx fˆ (ω)dω,

where the convergence is with respect to the norm in L2 (R).

Proof: Let f ∈ L2 (R). If g = fˆ , then, by Lemma 5.11.17,  n 1 f (x) = gˆ (x) = lim √ e−iωx g(ω)dω n→∞ 2π −n  n 1 = lim √ eiωx g(ω)dω n→∞ 2π −n  n 1 = lim √ eiωx fˆ (ω)dω. n→∞ 2π −n




Corollary 5.11.19. If f ∈ L1 (R) ∩ L2 (R), then the equality 1 f (x) = √ 2π



∞ −∞

eiωx fˆ (ω)dω

(5.95)

holds almost everywhere in R. The transform defined by (5.95) is called the inverse Fourier transform. One of the main reasons for introducing the factor √12π in the definition of the Fourier transform is the symmetry of the transform and its inverse:  ∞
 1 F f (x) = √ e−iωx f (x)dx, 2π −∞  ∞
 1 −1 F f (ω) = √ eiωx f (ω)dω. 2π −∞

Corollary 5.11.20. (Duality) If f ∈ L1 (R) ∩ L2 (R), then F {F {f (x)}} = f (−x)

almost everywhere in R.

Proof: It suffices to replace x by −x in (5.95). Theorem 5.11.21. (General Parseval’s relation) If f , g ∈ L2 (R), then 

∞ −∞

 f (x)g(x)dx =

∞ −∞

fˆ (ω)ˆg (ω)dω.




268

Chapter 5

Integral and Differential Equations

Proof: The polarization identity f , g =

 1 |f + g|2 − |f − g|2 + i|f + ig|2 − i|f − ig|2 4

implies that every isometry preserves inner product. Since the Fourier transform is an isometry on L2 (R), we have f , g = fˆ , gˆ .
The following theorem summarizes the results of this section. It is known as the Plancherel theorem (Michel Plancherel (1885–1967)).

Theorem 5.11.22. (Plancherel theorem) For every f ∈ L2 (R), there exists fˆ ∈

L2 (R) such that the following holds:

 ∞ 1 then fˆ (ω) = √ e−iωx f (x)dx, 2π −∞    n   1 ˆ (b)  e−iωx f (x)dx f (ω) − √2π  → 0 as n → ∞, −n 2    n   1 eiωx fˆ (ω)dω (c)   → 0 as n → ∞, f (x) − √2π −n 2  2 2 (d) f 2 = ˆf 2 , (a)

If f ∈ L1 (R) ∩ L2 (R),

(e)

The map f → fˆ is a Hilbert space isomorphism of L2 (R) onto L2 (R).

Proof: The only part of this theorem that remains to be proved is the fact that the Fourier transform is “onto.” Let f ∈ L2 (R), and define h=f

and

ˆ g = h.

Then, by Lemma 5.11.17, f = h = gˆ , and hence f = gˆ . This shows that every square integrable function is the Fourier transform of a square integrable function.


Theorem 5.11.23. The Fourier transform is a unitary operator on L2 (R). Proof: First note that, for any g ∈ L2 (R), we have 1 F {g}(ω) = √ 2π



∞

1 e−iωx g(x)dx = √ 2π −∞



∞ −∞

eiωx g(x)dx = F −1 {g}(ω).

Now, using Theorem 5.11.16, we obtain  ∞ % $ F {f }, g = F {f }(x)g(x)dx −∞

5.11 The Fourier Transform

 = =

∞

−∞  ∞ −∞

269

f (x)F {g}(x)dx
 f (x)F −1 g(x) dx

% $ = f , F −1 {g} . This shows that F −1 = F ∗ , and thus F is unitary.




The Fourier transform can be defined for functions in L1 (RN ) by  1 ˆf (ω) = e−iω·x f (x)dx (2π )N/2 RN where ω = (ω1 , . . . , ωN ), x = (x1 , . . . , xN ), and ω · x = ω1 x1 + · · ·+ ωN xN . The theory of the Fourier transform in L1 (RN ) is similar to the one-dimensional case. Moreover, the extension to L2 (RN ) is possible, and it has similar properties, including the inversion theorem and the Plancherel theorem. We close this section with some examples of Fourier transforms. We chose these particular functions because of their connection with the wavelet transform (see Chapter 8).

Example 5.11.24. (Second derivative of a Gaussian function) If  1 2  f (x) = 1 − x2 e− 2 x , then 1 2 fˆ (ω) = ω2 e− 2 ω .

Indeed, we have


 1 2 F 1 − x e− 2 x = −F 2



d2 − 1 x2 e 2 dx2
1 2 = −(iω)2 F e− 2 x ω2

= ω 2 e− 2 . The functions f (x) and fˆ (ω) are plotted in Figure 5.3.

Example 5.11.25. (The Haar function) The Haar function is defined by  1 if 0 ≤ x < 12 , f (x) = −1 if 1 ≤ x < 1, 2  0 otherwise.




270

Chapter 5

Integral and Differential Equations

Figure 5.3 Second derivative of a Gaussian function and its Fourier transform.

Clearly, 1 fˆ (ω) = √ 2π



1/2

e

−iωx

 dx −

0

1

e

−iωx

 dx

1/2

 iω 1 1 1 − 2e− 2 + e−iω =√ iω 2π iω  1 e− 2  iω =√ e 2 − 2 + e− 2 2π iω iω

=√

i 2π

e− 2

iω

sin2 ( ω4 ) ω 4

.

The Haar function is named after Alfréd Haar (1885–1933). The Haar function and its Fourier transform are drawn in Figure 8.2 in Chapter 8.


Example 5.11.26. (The Shannon function) The Shannon function is defined by f (x) =

sin 2π x − sin π x . πx

The Fourier transform of f (x) is fˆ (ω) =



√1 2π

0

if π < |ω| < 2π , otherwise.

The graph of the Shannon function and its Fourier transform are shown on Figure 5.4. The Shannon function is named after Claude Elwood Shannon (1916–2001).


5.12 Applications of the Fourier Transform

271

Figure 5.4 The Shannon function and its Fourier transform.

5.12 Applications of the Fourier Transform to Ordinary Differential Equations and Integral Equations In this section, we discuss some examples of applications of the Fourier transform to ordinary differential equations and integral equations. Consider the nth order linear ordinary differential equation with constant coefficients Ly(t) = f (t),

(5.96)

where L is the nth order differential operator given by L = an Dn + an−1 Dn−1 + · · · + a1 D + a0 ,

(5.97)

a0 , a1 , . . . , an are constants, D = dtd , and f ∈ L1 (R) or f ∈ L2 (R). Application of the Fourier transform to both sides of (5.96) gives  an (iω)n + an−1 (iω)n−1 + · · · + a1 (iω) + a0 yˆ (ω) = fˆ (ω)

 or

pˆ (iω)ˆy(ω) = fˆ (ω), where pˆ (z) = an zn + an−1 zn−1 + · · · + a1 z + a0 . Thus, yˆ (ω) =

fˆ (ω) = fˆ (ω)ˆg (ω), pˆ (iω)

where gˆ (ω) =

1 . pˆ (iω)

(5.98)

272

Chapter 5

Integral and Differential Equations

Now, the convolution Theorem 5.11.11 gives the solution  ∞ 1 y(t) = √ f (ξ )g(t − ξ )dξ, 2π −∞

(5.99)

provided g(t) = F −1 {ˆg (ω)} is known explicitly. To give a physical interpretation of the result, we consider the differential equation associated with a sudden impulse function f (t) = δ(t): LG(t) = δ(t).

(5.100)

(For a rigorous discussion of the Dirac delta distribution δ, see Section 6.2.) Application of the Fourier transform to (5.100) yields the solution  1 1 G(t) = F −1 √ gˆ (ω) = √ g(t). (5.101) 2π 2π Now, the solution (5.99) can be written as  ∞ f (ξ )G(t − ξ )dξ. y(t) = −∞

(5.102)

Clearly, G(t) behaves like a Green’s function, that is, it is the response to a unit impulse. In any physical system, f (t) is usually called the input function, while y(t) is called the √ output obtained by the superposition principle. The Fourier transform of 2πG(t) is called the admittance gˆ (ω) = (pˆ (iω))−1 . To determine the response to a given input, we first find the Fourier transform of the input, multiply the result by the admittance, and then apply the inverse Fourier transform to the product. We illustrate these ideas by a simple electrical circuit problem.

Example 5.12.1. The electric current I(t) in the circuit is governed by the equation dI + RI = E, (5.103) dt where L is the inductance, R is the resistance, and E is the applied electromagnetic force. With E(t) = E0 e−|t| , application of the Fourier transform (with respect to t) to Equation (5.103) gives . 2 E0 (iωL + R)Iˆ(ω) = π 1 + ω2 L

or

. Iˆ(ω) =

E0 2 . π (iωL + R)(1 + ω2 )

273

5.12 Applications of the Fourier Transform

The inverse Fourier transform yields  eiωt dω E0 ∞ . I(t) = π −∞ (iωL + R)(1 + ω2 ) This integral can readily be evaluated by the theory of residues. For t > 0,  , - E0 iR I(t) = 2π i [residue at ω = i] + residue at ω = π L   e−t e−Rt/L = 2iE0 + 2i(R − L) iL(1 − R2 /L2 )  −t  2Le−Rt/L e − 2 . = E0 R−L R − L2 Similarly, for t < 0, we obtain I(t) = −

E0 et E0 · 2π i[residue at ω = −i] = . π L+R

(5.104)

At t = 0, the current is continuous, hence I(0) = lim I(t) = t→0

E0 . R+L




Example 5.12.2. (Synthesis and resolution of a pulse — physical interpretation of convolution) A time-dependent electric, optical, or electromagnetic pulse can be regarded as a superposition of plane waves of all real frequencies so that the total pulse can be represented by the inverse Fourier transform  ∞ 1 f (t) = F(ω)eiωt dω, (5.105) 2π −∞ where the factor 1/2π is introduced because the angular frequency ω is related to the linear frequency ν by ω = 2π ν, and negative frequencies are introduced for mathematical convenience so that we can avoid dealing with the cosine and sine functions separately. Clearly, F(ω) can be represented by the Fourier transform of f (t) as  ∞ F(ω) = f (t)e−iωt dt. (5.106) −∞

This represents the resolution of the pulse f (t) into its angular frequency components, and (5.105) gives a synthesis of the pulse from its individual components. Consider a simple electrical device such as amplifier with an input function f (t) and an output function g(t). For an input of a single frequency ω, we have

274

Chapter 5

Integral and Differential Equations

f (t) = eiωt . The amplifier will change the amplitude and may also change the phase so that the output can be expressed in terms of the input, the amplitude, and phase modifying function (ω) as g(t) = (ω)f (t),

(5.107)

where (ω) is usually called the transfer function, and it is, in general, a complex function of the real variable ω. This function is generally independent of the absence or presence of any other frequency components. Thus, the total output may be obtained by integrating over the entire input as modified by the amplifier  ∞ 1 g(t) = (ω)F(ω)eiωt dω. (5.108) 2π −∞ Therefore, the total output g(t) can readily be calculated from any given input f (t) and known transfer function (ω). On the other hand, the transfer function is obviously characteristic of the amplifier and can, in general, be obtained as the Fourier transform of some function ϕ(t):  ∞ (ω) = ϕ(t)e−iωt dt. (5.109) −∞

The convolution Theorem 5.11.11 allows us to rewrite (5.108) as  ∞
 f (τ )ϕ(t − τ )dτ. g(t) = F −1 (ω)F(ω) = −∞

(5.110)

Physically, the result represents an output signal (effect) as the integral superposition of an input signal (cause) f (t) modified by ϕ(t − τ ). Indeed, (5.110) is the most general mathematical representation of an output in terms of an input modified by the amplifier where t is the time variable. Assuming the principle of causality, that is, every effect has a cause, we must require τ < t. The principle of causality is imposed by requiring ϕ(t − τ ) = 0 for τ > t. Consequently, (5.110) reduces to the form  t g(t) = f (τ )ϕ(t − τ )dτ. −∞

(5.111)

(5.112)

To determine the significance of ϕ(t), we use a sudden impulse function f (τ ) = δ(τ ) so that (5.112) becomes  t g(t) = δ(τ )ϕ(t − τ )dτ = ϕ(t)H(t). (5.113) −∞

5.12 Applications of the Fourier Transform

275

This recognizes ϕ(t) as the output corresponding to a unit impulse at t = 0, and the Fourier transform (ω) of ϕ(t) is given by  ∞ (ω) = ϕ(t)e−iωt dt (5.114) 0

with ϕ(t) = 0 for t < 0.




Example 5.12.3. We will use the Fourier transform to solve the following ordinary differential equation: −

d2 u + α 2 u = f (x), dx2

(5.115)

where f ∈ L2 (R). Applying the Fourier transform to both sides of (5.115), we obtain  2  ω + α 2 uˆ (ω) = fˆ (ω). Hence, uˆ (ω) = Since 1 √ 2π we have



∞

−∞

ω2

1 fˆ (ω). + α2

1 −α|x| eiωx dω = , e 2 2 ω +α 2α

 1 1 −α|x| = F . e ω2 + α 2 2α

From the convolution Theorem 5.11.11, we get   1 −α|x| 1 −α|x| F {u} = F F {f } = F ∗f e e 2α 2α and hence

 u(x) =

1 −α|x| ∗f. e 2α

This means that the solution of the equation can be written as  ∞ −α|x−t| e 1 f (t)dt. u(x) = √ 2π −∞ 2α




Example 5.12.4. Consider an infinite beam on an elastic foundation under a prescribed vertical load W(x). The vertical deflection u(x) is governed by the

276

Chapter 5

Integral and Differential Equations

following differential equation EIu(iv) + ku = W(x),

(5.116)

where EI is the flexural rigidity, and k is the foundation modulus of the beam. We solve this equation assuming that W(x) has compact support so that u, u , u

, and u

all tend to zero as |x| → ∞. We first rewrite the equation: u(iv) + α 4 u = w(x),

(5.117)

where α 4 = k/EI, and w(x) = W(x)/EI. Application of the Fourier transform to (5.117) gives w(ω) ˆ . 4 ω + α4 The inverse Fourier transform yields the solution  ∞ 1 w(ω) ˆ u(x) = √ eiωx dω 4 + α4 ω 2π −∞  ∞  ∞ iωx e dω 1 = w(ξ )e−iωξ dξ 2π −∞ ω4 + α 4 −∞  ∞ = w(ξ )G(ξ, x)dξ, uˆ (ω) =

−∞

where 1 G(ξ, x) = 2π



∞

eiω(x−ξ ) 1 dω = 4 + α4 ω π −∞



∞ 0

cos(x − ξ )ω dω. ω4 + α 4

(5.118) (5.119) (5.120)

(5.121)

This integral can be evaluated by complex contour integration (or with the aid of tables of Fourier integrals). We simply state the result:   1 −α|x−ξ |/√2 α|x − ξ | π G(ξ, x) = 3 e . (5.122) sin + √ 2α 4 2 In particular, consider a point load of unit strength acting at some point x0 , that is, w(x) = δ(x − x0 ). Then the solution (5.120) becomes  ∞ u(x) = δ(ξ − x0 )G(ξ, x)dξ = G(x0 , x). (5.123) −∞

Thus, the kernel G(ξ, x) in (5.120) has the physical significance of being the deflection, as a function of x, due to a point unit load at ξ . So the deflection due to a point load of strength w(ξ )dξ at ξ is w(ξ )dξ G(ξ, x), and hence (5.120) represents the superposition of all such incremental deflections.


277

5.12 Applications of the Fourier Transform

Example 5.12.5. We will solve the integral equation of the convolution type: 

∞

−∞

K(x − t)u(t)dt + λu(x) = f (x),

where K, f ∈ L1 (R). This is a Fredholm integral equation with convolution kernel. Application of the Fourier transform gives √ ˆ 2π K(ω)ˆ u(ω) + λˆu(ω) = fˆ (ω). The inverse Fourier transform leads to a formal solution in the form  ∞ fˆ (ω)eiωx 1 u(x) = √ dω. (5.124) √ ˆ 2π −∞ 2π K(ω) +λ As an example, consider K(x) = 1/x. Then . .  ∞ −iωx 1 e π π ˆ dx = i sgn(−ω) = −i sgn ω. K(ω) =√ 2 2 2π −∞ x Thus, the solution (5.124) becomes  ∞ ˆ f (ω)eiωx 1 dω. u(x) = √ 2π −∞ λ − iπ sgn ω




Example 5.12.6. (The Hilbert transform) We solve the integral equation 1 π



∞

f (t) dt = fH (x), x −∞ − t

(5.125)

where fH ∈ L2 (R) and the above integral is the Cauchy principal value:  −ε  ∞   ∞ f (t) f (t) dt = lim dt. + ε→0 x−t −∞ x − t −∞ ε The function fH (x) is called the Hilbert transform of f (t) and is often denoted by H{f (t)} = fH (x). Our problem is thus to find the inverse Hilbert transform. First we rewrite Equation (5.125) as  ∞ 1 f (t)g(x − t)dt = fH (x), √ 2π −∞ √ where g(x) = 2/π (1/x). Application of the Fourier transform with respect to x gives fˆH (ω) fˆ (ω) = , gˆ (ω)

gˆ (ω) = −i sgn ω.

278

Chapter 5

Integral and Differential Equations

The inverse Fourier transform yields the solution:  ∞ 1 f (x) = − √ (i sgn ω)fˆH (ω)eiωx dω. 2π −∞ By convolution theorem, this reduces to the form  
1 ∞ fH (ξ ) dξ = −H fH (ξ ) . f (x) = − π −∞ x − ξ Consequently, H2 f = −f and thus H−1 = −H.

(5.126)


Example 5.12.7. Consider the integral equation 

∞

1 e−|x−t| u(t)dt = − u(x) + e−|x| , 4 −∞

−∞ < x < ∞.

(5.127)

Application of the Fourier transform with respect to x yields . 2 1 2 1 uˆ (ω) = − uˆ (ω) + , 2 1+ω 4 π 1 + ω2 so that 1 8 uˆ (ω) = √ . 2+9 ω 2π By the inverse Fourier transform, we obtain  4 ∞ eiωx dω. u(x) = π −∞ ω2 + 9 To evaluate this integral for x > 0, we use a semicircular closed contour in the lower half of the complex plane. It turns out that 4 u(x) = e−3x , 3

x > 0.

Similarly, for x < 0, we use a closed semicircular contour in the upper half of the complex plane to obtain 4 u(x) = e3x , 3

x < 0.

Hence, the solution of (5.127) is 4 u(x) = e−3|x| . 3

(5.128)


5.13 Exercises

279

5.13 Exercises 1. Determine the fixed points, if any, of the following operators: (a) T(x) = x + a on any vector space. (b) T : R2 → R2 defined by T(x, y) = (x, 0). (c) T : R2 → R2 defined by T(x, y) = (y, y). (d) T : R2 → R2 defined by T(x, y) = (x cos ϕ + y sin ϕ, −x sin ϕ + y cos ϕ), where ϕ is a fixed real number. 2. Suppose T is an operator on C ([0, 1]) defined by  t  2 (Tu)(t) = u(x) dx. 0

Show that T is not a contraction on the closed unit ball in C ([0, 1]), but that it is one on the closed ball of radius 14 in C ([0, 1]). 3. Show that the operator T : C ([0, 1]) → C ([0, 1]) defined by  t (Tx)(t) = x(0) + λ x(τ )dτ, λ ∈ R, 0

is a contraction provided |λ| < 1. 4. Complete the proof of Theorem 5.2.2 by showing that x0 = limn→∞ T n x for any x ∈ E. 5. Show that the nonlinear integral equation  1   f (x) = e−sx cos αf (s) ds, 0 ≤ x ≤ 1, 0 < α < 1, 0

has a unique solution. 6. Consider a system of ordinary differential equations   d ϕk (x) = fk x0 , ϕ1 (x), ϕ2 (x), . . . , ϕN (x) dx with the initial data ϕk (x0 ) = y0k , where k = 1, 2, . . . , N; and the functions fk (x0 , y1 , y2 , . . . , yN ) are continuous in some domain  ⊂ RN+1 , and (x0 , y01 , y02 , . . . , y0N ) ∈ .

280

Chapter 5

Integral and Differential Equations

Moreover, we assume that the functions fk satisfy the Lipschitz condition   fk (x, y1 , y2 , . . . , yN ) − fk (x, z1 , z2 , . . . , zN ) ≤ sup max |ym − zm | x 1≤m≤N

in . Prove that this system has a unique system of solutions yk = ϕk (x) in some interval |x − x0 | < d. 7. Use the method presented in Section 5.6 to solve the following homogeneous Fredholm equation:  1 f (x) = λ (x + t)f (t)dt. −1

8. Use the method presented in Section 5.6 to solve the following nonhomogeneous equation:  1   f (x) = ϕ(x) + λ π x sin π t + 2π x2 sin 2π t f (t)dt. 0

9. Express the solution of the integral equation  2π f (x) = ϕ(x) + λ cos(x + t)f (t)dt 0

in the resolvent form  f (x) = ϕ(x) + λ

2π

(x, t; λ)ϕ(t)dt, 0

where λ is not an eigenvalue. Obtain the general solution, if it exists, for ϕ(x) = sin x. 10. Show that the solution of the differential equation d2 f + xf = 1, dx2

f (0) = f (0) = 0,

satisfies the nonhomogeneous Volterra equation  x x2 t(t − x)f (t)dt. f (x) = + 2 0 11. Transform the problems (a)

d2 f + f = x, dx2

f (0) = 0,

f (1) = 0,

5.13 Exercises

d2 f + f = x, f (0) = 1, dx2 into Fredholm integral equations.

(b)

281

f (1) = 0,

12. Discuss the solutions of the integral equation  1 (x + t)f (t)dt. f (x) = ϕ(x) + λ 0

13. When do the following integral equations have solutions ? 1 (a) f (x) = ϕ(x) + λ 0 (1 − 3xt)f (t)dt.  2π (b) f (x) = ϕ(x) + λ 0 sin(x + t)f (t)dt. 1 (c) f (x) = ϕ(x) + λ 0 xtf (t)dt. 1  (d) f (x) = ϕ(x) + λ −1 m n=1 Pn (x)Pn (t)f (t)dt, where Pn is the nth degree Legendre polynomial. 1 (e) f (x) = x + 12 −1 (x + t)f (t)dt. 14. Find the eigenvalues and eigenfunctions of the following integral equations:  2π (a) f (x) = λ 0 cos(x − t)f (t)dt. 1 (b) f (x) = λ −1 (t − x)f (t)dt.  2π (c) f (x) = ϕ(x) + λ 0 cos(x + t)f (t)dt. 15. Solve the integral equations 1 (a) f (x) = ϕ(x) + λ 0 tf (t)dt.  1/2 (b) f (x) = x + λ 0 f (t)dt.  1 1 (c) f (x) = 5x 6 + 2 0 xtf (t)dt. 1 (d) f (x) = x + 0 (1 + xt)f (t)dt. 1 (e) f (x) = ex + λ 0 2ex+t f (t)dt. 16. Use the separable kernel method to show that  1 f (x) = λ cos x sin tf (t)dt 0

has no solution except the trivial solution f = 0. 17. Obtain the Neumann series solutions of the following equations:

282

Chapter 5

Integral and Differential Equations

 1 1 2 −1 (t + x)f (t)dt. x f (x) = x + 0 (t − x)f (t)dt. x f (x) = x − 0 (t − x)f (t)dt. x f (x) = 1 − 2 0 tf (t)dt.

2

(a) f (x) = x + (b) (c) (d)

18. If Lu = u + ω u, show that L is formally self-adjoint and the concomitant is J(u, v) = vu − uv . Moreover, if u is a solution of Lu = 0 and v is a solution of L∗ v = 0, then the concomitant of u and v is a constant. 19. Let L be a self-adjoint differential operator given by (5.64). If u1 and u2 are two solutions of Lu = 0, and J(u1 , u2 ) = 0 for some x for which a2 (x) = 0, then show that u1 and u2 are linearly independent. 20. Consider the differential operator d , 0 ≤ x ≤ 1, dx u(1) = 0.

L = ex D2 + ex D, u (0) = 0,

D=

Show that L is formally self-adjoint. 21. Prove continuity of the Green’s function defined in Theorem 5.10.1. 22. Find eigenvalues and eigenfunctions of the following Sturm–Liouville system: u

+ λu = 0,

0 ≤ x ≤ π,

u(0) = u (π ) = 0. 23. Transform the Euler equation x2 u

+ xu + λu = 0,

1 ≤ x ≤ e,

with the boundary conditions u(1) = u(e) = 0 into the Sturm–Liouville system , d du 1 x + λu = 0, dx dx x u(1) = u(e) = 0. Find the eigenvalues and eigenfunctions.

5.13 Exercises

283

24. Prove that λ = 0 is not an eigenvalue of the system defined in Example 5.9.3. 25. Show that the Sturm–Liouville operator L = DpD + q, D = d/dx, is positive if p(x) > 0 and q(x) ≥ 0 for all x ∈ [a, b]. 26. Show that the Sturm–Liouville operator L in L2 ([a, b]) given by

L=

1 (DpD + q) r(x)

is not symmetric. 27. Provide a detailed proof for Corollary 5.11.19. 28. Prove Theorem 5.11.7. 29. Find the Fourier transform of  1 if x ∈ [−a, a], (a) f (x) = 0 otherwise,  |x| (b) f (x) = 1 − 2 if x ∈ [−2, 2], 0 otherwise,   1 2 (c) f (x) = 1 − x2 e− 2 x ,  ∞ (d) f (x) = g(x + t)g(t)dt. −∞

30. Show that under appropriate conditions (a) fˆ (ω) = −iF {xf (x)}, (b) fˆ (r) (ω) = (−i)r F {xr f (x)}. 31. Use the Parseval relation to evaluate   ∞ sin x 2 (a) dx, x −∞   ∞ sin x 3 dx, (b) x −∞   ∞ sin x 4 dx. (c) x −∞ 32. Show that 9 f (ω) ∗9 g(ω) = F {f (x)g(x)}. 33. Show that the eigenvalues of the Fourier transform are λ = 1, −1, i, −i.

284

Chapter 5

Integral and Differential Equations

34. The Hermite functions are given by hn (x) =

1 −x2 /2 e Hn (x), n!

where Hn (x) is the Hermite polynomial of degree n (see Example 5.8.15). Show that 9 hn (ω) = (−i)n hn (ω). 35. Use the Fourier transform to solve the forced linear harmonic oscillator x¨ + ω2 x = a sin t,

t > 0, ω = ,

x(0+) = 0 = x˙ (0+).

Examine the case when ω = . 36. Solve the problem discussed in Example 5.12.1 with E(t) = E0 e−αt sin ωt · H(t) and I(0+) = I0 . 37. If there is a capacitor in the circuit discussed in Example 5.12.1, then the current I(t) satisfies the following integro-differential equation: ,  t dI 1 I(t)dt = E(t), L + RI + q0 + dt C 0 where q0 is the initial charge on the capacitor so that  t I(t)dt q = q0 + 0

is the charge and dq/dt = I. Solve this problem using the Fourier transform and the following conditions: I = q = E = 0 for t < 0, I(0+) = I0

and

q(0+) = q0 .

Examine the special case when E(t) = H(t). 38. Use the Fourier transform to solve the following problem: y

+ 3y + 2y = e−x ,

x > 0,

y(0+) = y0

and

y (0+) = y00 .

39. Use the Fourier transform to solve the following pair of coupled differential systems for t > 0: x + y − x + 3y = e−t ,

5.13 Exercises

285

x + y + 2x + y = e−2t , x(0+) = x0

and

y(0+) = y0 .

40. Show that the formal solution of x¨ + 2k˙x + σ 2 x = f (t) is



1 x(t) = √ 2π

∞ −∞

fˆ (ω)eiωt dω . σ 2 − ω2 + 2iωk

41. (a) Show that the solution of the integral equation  ∞ u(x) − λ e−|x−t| u(t)dt = e−|x| −∞

is

√ 1−2λ|x|

e− u(x) = √ (b) If

 Tu =

1 − 2λ ∞

−∞

1 for λ > . 2

e−|x−t| u(t)dt,

show that T ≤ 2; · denotes the norm in L2 (R). 42. Prove the following properties of the Hilbert transform ϕ(x) ˜ = H{ϕ(t); x}: (a) H{ϕ(t + a); x} = H{ϕ(t); x + a}. (b) H{ϕ(at); x} = H{ϕ(t); ax}, a > 0. (c) H{ϕ(−t); x} = −H{ϕ(t); −x}. (d) H{ϕ (t); x} = (d/dx)ϕ(x). ˜ ˜ + (1/π ) (e) H{tϕ(t); x} = xϕ(x)

∞

−∞ ϕ(t)dt.

43. Show that if f ∈ L2 (R), then H{f } ∈ L2 (R) and H{f } = f . 44. Show that F {H{f }} = (−i sgn ω)F {f }.

This page intentionally left blank

Generalized Functions and Partial Differential Equations

CHAPTER

6

“However varied may be the imagination of man, nature is still a thousand times richer. . . . Each of the theories of physics . . . presents (partial differential) equations under a new aspect . . . without these theories, we should not know partial differential equations.” Henri Poincaré “Between 1930 and 1940 several mathematicians began to investigate systematically the concept of a “weak” solution of a linear partial differential equation, which appeared episodically (and without a name) in Poincaré’s work . . . . It was one of the main contributions of Laurent Schwartz when he saw, in 1945, that the concept of distribution introduced by Sobolev (which he had rediscovered independently) could give a satisfactory generalization of the Fourier transform including all the preceding ones.” Jean Dieudonné

6.1 Introduction In this chapter, we shall first discuss briefly the basic concepts and properties of distributions. The theory of distributions was initiated by S.L. Sobolev in 1936. The concept of distributions was independently introduced by Laurent Schwartz (1915–2002) in the 1950s. Since Schwartz was the one who developed the theory almost to its present form, distributions are often called Schwartz distributions. Distributions have found applications in many areas of mathematics and physics, including differential and integral equations.

287

288

Chapter 6

Partial Differential Equations

In Section 6.3, we discuss Sobolev spaces. The definitions and some basic properties are presented. The remainder of this chapter deals with Green’s functions for partial differential equations of most common interest. This is followed by the form of the Green’s identity associated with partial differential operators. Section 6.5 discusses weak solutions of elliptic boundary value problems. The final section is devoted to examples of applications of the Fourier transform to partial differential equations of physical interest.

6.2 Distributions Consider a partial differential operator L of order m in N variables  aα Dα , L=

(6.1)

|α|≤m

where α = (α1 , . . . , αN ) is a multi-index, αn ’s are non-negative integers, |α| = α1 + · · · + αN , aα = aα1 ,α2 ,...,αN are constant coefficients, and     ∂ α1 ∂ αN ∂ |α| α ... = α1 D = αN . ∂x1 ∂xN ∂x1 . . . ∂xN Let f ∈ C (RN ) and let u be a function satisfying the equation  aα Dα u = f . Lu =

(6.2)

|α|≤m

This equation makes sense only if all derivatives of u appearing in L exist. This puts a rather severe restriction on the family of solutions and, as we will see later, quite often there is no such solution. Let ϕ be an infinitely differentiable function on RN vanishing outside of some bounded set. First we multiply both sides of Equation (6.2) by ϕ(x) to obtain    aα Dα u(x) ϕ(x) = f (x)ϕ(x). (6.3) |α|≤m

Next we integrate (6.3) over RN :       α aα D u(x) ϕ(x) dx = RN

|α|≤m

RN

f (x)ϕ(x)dx.

(6.4)

Since ϕ(x) = 0 outside a bounded subset of RN , integration by parts yields     |α| α aα u(x)(−1) D ϕ(x) dx = f (x)ϕ(x)dx. (6.5) RN

|α|≤m

RN

6.2 Distributions

289

If we put 

L∗ =

(−1)|α| aα Dα ,

(6.6)

|α|≤m

then (6.5) can be written as 

  u(x) L∗ ϕ (x)dx =

RN

 RN

f (x)ϕ(x)dx.

(6.7)

This analysis shows that every function u satisfying (6.2) also satisfies (6.7). On the other hand, a function u may satisfy (6.7) without being differentiable. This observation becomes the fundamental idea behind the theory of generalized functions. Let f and g be continuous functions on RN . The condition f (x) = g(x) for every x ∈ RN follows if



 RN

f (x)ϕ(x)dx =

RN

g(x)ϕ(x)dx

for every ϕ ∈ C ∞ (RN ) vanishing outside of some bounded set. The essential difference is that instead of comparing the values at every point x ∈ RN , we test functions f and g using smooth functions with bounded support. For this reason, those functions are called test functions. They play the crucial role in the theory of distributions.

Definition 6.2.1. (Test function) By a test function, we mean an infinitely differentiable function on RN vanishing outside of some bounded set. The space of all test functions is denoted by D(RN ) or simply by D . Instead of an infinitely differentiable function, we will often say a smooth function.

Example 6.2.2. The existence of nontrivial test functions is not obvious. In the case when the number of variables N = 1, an example of a test function is 

(x ϕ(x) = e 0

2 −1)−1

if |x| < 1, otherwise

290

Chapter 6

Partial Differential Equations

−1

Figure 6.1 The graph of ϕ(x) = e(x −1) for |x| < 1. 2

(see Figure 6.1). Using this test function, we can easily generate a number of examples. The following are test functions: ϕ(ax + b), where a, b are constants, a = 0, f (x)ϕ(x),

where f is an arbitrary smooth function,

(k)

where k is a positive integer.

ϕ (x), The function



(x1 +···+xN −1) ϕ(x) = e 0 2

2

−1

2 if x12 + · · · + xN < 1, otherwise

is a test function on RN . Another way to obtain examples of test functions on RN is to take arbitrary test functions ϕ1 , . . . , ϕN defined on R and then define ϕ(x) = ϕ1 (x1 )ϕ2 (x2 ) · · · ϕN (xN ).




Theorem 6.2.3. The space of test functions D is a vector space. Moreover, if

ϕ, ψ ∈ D , then

(a) f ϕ ∈ D for every smooth function f , (b) ϕ ◦ A ∈ D for every affine transformation A of RN onto RN , (c) ϕ ∗ ψ ∈ D .




The proof is left as an exercise. To define distributions, we need to introduce a convergence in the space of test functions.

Definition 6.2.4.

(Convergence of test functions) Let ϕ1 , ϕ2 , . . . and ϕ be test functions. We say that the sequence (ϕn ) converges to D ϕ in D , denoted by ϕn −→ ϕ , if the following two conditions are satisfied: (a) ϕ1 , ϕ2 , . . . and ϕ vanish outside some bounded set S ⊂ RN ,

6.2 Distributions

291

(b) Dα ϕn → Dα ϕ uniformly on RN for every multi-index α .

Example 6.2.5. Let ϕ ∈ D and let {vn } be a sequence of vectors in RN converD

gent to 0. Define ϕn (x) = ϕ(x − vn ). Then ϕn −→ ϕ. In other words, translation is a continuous operation in D . Let ϕ ∈ D , and let (an ) be a sequence of scalars convergent to a scalar a. D

Then an ϕ −→ aϕ. In other words, multiplication by scalars is a continuous operation on D . Let ϕ be a nonzero test function, and let (an ) and (bn ) be sequences of positive scalars convergent to 0. Define ϕn (x) = bn ϕ(an x). Then Dα ϕn → 0 uniformly on RN for every multi-index α. In this case, however, the sequence {ϕn } is not convergent in D because condition (a) is not satisfied. The supports of ϕn ’s expand without a bound as n → ∞.
The properties of the convergence in D listed in the following theorem are direct consequences of the definition. The easy proofs are left as an exercise. D

D

Theorem 6.2.6. Let ϕn −→ ϕ and ψn −→ ψ. Then D

(a) aϕn + bψn −→ aϕ + bψ, for any scalars a, b, D

(b) f ϕn −→ f ϕ, for any smooth function f defined on RN , D

(c) ϕn ◦ A −→ ϕ ◦ A, for any affine transformation A of RN onto RN , D

(d) Dα ϕn −→ Dα ϕ, for any multi-index α.




Definition 6.2.7.

(Distribution) By a distribution F on RN , we mean a continuous linear functional on D(RN ). In other words, a mapping F : D(RN ) → C is called a distribution if (a) F(aϕ + bψ) = aF(ϕ) + bF(ψ) for every a, b ∈ C and ϕ, ψ ∈ D(RN ), D

(b) F(ϕn ) → F(ϕ) (in C) whenever ϕn −→ ϕ . The space of all distributions is denoted by D (RN ) or simply by D . It will be convenient to write F, ϕ instead of F(ϕ). Note that F, ϕ is not an inner product. Since it is always clear from the context what is meant, this inconsistency does not lead to any problems.

292

Chapter 6

Partial Differential Equations

Distributions generalize the concept of a function. Formally, a function on RN is not a distribution because its domain is not D . However, every locally integrable function f on RN can be identified with a distribution F in the following way:  F, ϕ = f ϕ. RN

Definition 6.2.8.

(Regular and singular distributions) A distribution F ∈ D is called a regular distribution if there exists a locally integrable function f such that  F, ϕ = f ϕ, (6.8) RN

for every ϕ ∈ D . A distribution which is not regular is called a singular distribution. The proof of the fact that (6.8) defines a distribution is left as an exercise. Note that there is no problem with integrability of the product f ϕ because it vanishes outside a bounded set. The integral RN f ϕ can be interpreted, at least for some test functions ϕ, as the average value of f with respect to probability whose density function is ϕ. Thus, one can think of F, ϕ as an average value of F, and of distributions as objects that have average values in the neighborhood of every point. In general, distributions do not have values at points. This interpretation of distributions is very natural from the point of view of physics: when a quantity is measured, the result is not the exact value at a single point.

Example 6.2.9. Let  be an open (or just measurable) set in RN . The functional F defined by

 F, ϕ =

ϕ 

is a distribution. Observe that it is a regular distribution since  F, ϕ = χ ϕ, RN

where χ is the characteristic function of . In particular, if  = (0, ∞) × · · · × (0, ∞), we obtain a distribution  ∞  ∞ ... ϕ(x)dx1 . . . dxN , H, ϕ = 0

0

which is called the Heaviside function (Oliver Heaviside (1850–1925)). The letter H will be used to denote this distribution, as well as the characteristic function of  = (0, ∞) × · · · × (0, ∞).


293

6.2 Distributions

Example 6.2.10. (Dirac delta distribution) One of the most important examples of distributions is the so called Dirac delta (Paul Adrien Maurice Dirac (1902–1984)). It is denoted by δ and defined by δ, ϕ = ϕ(0).

(6.9) D

The linearity of δ is obvious. To prove continuity note that ϕn −→ ϕ implies ϕn → ϕ uniformly on RN , and thus ϕn (x) → ϕ(x) for every x ∈ RN . The Dirac delta is a singular distribution.


Example 6.2.11. Let α be a multi-index. The functional F on D defined by F, ϕ = Dα ϕ(0)

(6.10)

is a distribution. The proof is similar to the proof of the fact that the Dirac delta is a distribution. Actually, distributions defined by (6.10) are closely related to δ (see Example 6.2.14).
The success of the theory of distributions is largely due to the fact that most concepts of calculus can be defined for distributions. When adopting definitions for distributions, we expect that new definitions agree with classical ones when applied to regular distributions. The following approach will ensure that. When looking for an extension of some operation A, that is defined for functions, we first consider regular distributions:  F, ϕ = f ϕ. Since we expect AF to be the same as Af , it is natural to define  AF, ϕ = Af ϕ. If we can find a continuous operation A∗ which maps D into D such that   Af ϕ = fA∗ ϕ, then it makes sense to define, for an arbitrary distribution F, % $ AF, ϕ = F, A∗ ϕ . For example, if the described method is used to find a natural definition of the derivative of a distribution, it suffices to note that   ∂ ∂ f (x)ϕ(x)dx = − f (x) ϕ(x)dx. ∂xk RN ∂xk RN

294

Chapter 6

Partial Differential Equations

Definition 6.2.12.

(Derivatives of distributions) Let F be a distribution. Then the derivative ∂F/∂xk is defined by 2 1 1 2 ∂ϕ ∂F , ϕ = − F, . ∂xk ∂xk

(6.11)

More generally, if α is a multi-index, then by Dα F we denote the functional defined by % % $ α $ (6.12) D F, ϕ = (−1)|α| F, Dα ϕ .

The fact that Dα F, ϕ is well defined for every ϕ ∈ D does not mean that D F is a distribution. That has to be proved. α

Theorem 6.2.13. If F is a distribution, then Dα F is a distribution for any multiindex α.

Proof: We need to show that Dα F is linear and continuous. If F is a distribution, then % $ % $ α D F, aϕ + bψ = F, a(−1)|α| Dα ϕ + b(−1)|α| Dα ψ $ % $ % = a F, (−1)|α| Dα ϕ + b F, (−1)|α| Dα ψ $ % $ % = a Dα F, ϕ + b Dα F, ψ , D

D

so that Dα F is linear. Moreover, since ϕn −→ ϕ implies Dα ϕn −→ Dα ϕ, the functional Dα F is continuous. Thus, Dα F is a distribution.


Example 6.2.14. Let H denote the Heaviside function of a single variable x. Then

1

2 1 2  ∞ d d ϕ (x)dx = ϕ(0) = δ, ϕ. H, ϕ = − H, ϕ = − dx dx 0

Therefore,

Moreover,

and similarly,

d H = δ. dx

(6.13)

2 % $ d δ, ϕ = δ , ϕ = −ϕ (0) dx

(6.14)

2 $ % dn δ, ϕ = δ (n) , ϕ = (−1)n ϕ (n) (0). n dx

(6.15)

1

1

295

6.2 Distributions

In general,

% $ α D δ, ϕ = (−1)|α| Dα ϕ(0).

(6.16)


Definition 6.2.15.

(Weak distributional convergence) A sequence of distributions (Fn ) is convergent to a distribution F if for every ϕ ∈ D.

Fn , ϕ → F, ϕ

This type of convergence is called the weak distributional convergence. Since this is the only convergence of distributions considered in this book, the notation Fn → F or limn→∞ Fn = F will be used.

Example 6.2.16. Let f1 , f2 , . . . be a sequence of continuous functions on RN .

Suppose, for some continuous function f , fn → f uniformly on every compact subset of RN . Define regular distributions  Fn , ϕ = fn ϕ, n = 1, 2, . . . , RN

and

 F, ϕ =

RN

f ϕ.

Then Fn → F. Indeed, if  denotes the support of a test function ϕ, then    fn ϕ = fn ϕ → f ϕ = F, ϕ. Fn , ϕ = RN








Example 6.2.17. Letf1 , f2 , . . . ∈ L1 (RN ). Suppose, for some f ∈ L1 (RN ), fn →

f in L1 (RN ), that is, RN |fn − f | → 0. Define regular distributions Fn and F as in the previous example. Then Fn → F. In fact, we have          Fn , ϕ−F, ϕ =  (fn −f )ϕ  ≤ |fn −f ||ϕ| ≤ sup |ϕ| |fn −f | → 0,  RN

for every ϕ ∈ D .

RN

RN

RN




Instead of saying “the sequence of regular distributions generated by a sequence of functions (fn ) converges to a distribution F,” we will say simply “the sequence of functions (fn ) converges to a distribution F in the distributional sense” or just “the sequence (fn ) is distributionally convergent to F.”

296

Chapter 6

Partial Differential Equations

Figure 6.2 Functions fn (x) =

n π(1+n2 x2 )

for n = 1, 2, 3.

Example 6.2.18. Consider the sequence of functions on R defined by fn (x) =

n , π(1 + n2 x2 )

n = 1, 2, . . .

(see Figure 6.2). We will show that the sequence (fn ) is distributionally convergent to the Dirac delta distribution δ. Let ϕ be a test function with the support contained in the interval [−a, a]. We need to show that  ∞ fn (x)ϕ(x)dx → ϕ(0) as n → ∞, −∞

or, equivalently, 

∞ −∞

Since



∞ −∞

fn (x)ϕ(x)dx − ϕ(0) → 0 as n → ∞. 

fn (x)dx =

∞

ndx = 1 for all n ∈ N, π(1 + n2 x2 ) −∞

we have  ∞       = f (x)ϕ(x)dx − ϕ(0) n    −∞

    ≤ ϕ(0)

−a

−∞

    fn (x)dx + 

∞ −∞

a −a

    fn (x) ϕ(x) − ϕ(0) dx

         fn (x) ϕ(x) − ϕ(0) dx + ϕ(0)

a

∞

  fn (x)dx.

297

6.2 Distributions

Direct integration shows that     −a        lim ϕ(0) fn (x)dx = lim ϕ(0) n→∞ n→∞ −∞

∞

a

  fn (x)dx = 0.

It thus remains to prove that   a      lim  fn (x) ϕ(x) − ϕ(0) dx = 0. n→∞

Note that

   

a −a

−a

(6.17)

  a       fn (x) ϕ(x) − ϕ(0) dx fn (x) ϕ(x) − ϕ(0) dx ≤ −a

  ≤ maxϕ (x)



a −a

 xfn (x)dx,

since |ϕ(x) − ϕ(0)| ≤ max |ϕ (x)||x|, by the mean value theorem, and  a 2 2   xfn (x)dx = lim ln(1 + n a ) = 0. lim n→∞ −a n→∞ πn




Theorem 6.2.19. If Fn → F in D (R), then Dα Fn → Dα F for every multi-index α.

Proof: We have $

% % % $ % $ $ Dα Fn , ϕ = (−1)|α| Fn , Dα ϕ → (−1)|α| F, Dα ϕ = Dα F, ϕ

for every test function ϕ.




Note that Theorem 6.2.19 allows us to differentiate convergent sequences or series of distributions term by term.

Definition 6.2.20.

(Antiderivative of a distribution) Let F ∈ D (R). A distribution G on R is called an antiderivative of F if G = F .

Theorem 6.2.21. Every distribution has an antiderivative. Proof: Let ϕ0 ∈ D(R) be a fixed test function such that 

∞ −∞

ϕ0 (x)dx = 1.

(6.18)

298

Chapter 6

Partial Differential Equations

Then, for every test function ϕ ∈ D (R), there exists a test function ϕ1 ∈ D(R) such that ϕ = Kϕ0 + ϕ1 , where

 K=



∞

ϕ(x)dx −∞

and

∞ −∞

ϕ1 (x)dx = 0.

Let F ∈ D (R). Define a functional G on D by G, ϕ = G, Kϕ0 + ϕ1  = KC0 − F, ψ,

(6.19)

where C0 is a constant and ψ is the test function defined by  x ϕ1 (t)dt. ψ(x) = −∞

Then G is a distribution and G = F.




It can be proved that if G1 and G2 are antiderivatives of a distribution F, then G1 − G2 is a constant function, or, more precisely, there exists a constant C such that  ∞ G1 − G2 , ϕ = C ϕ(x)dx −∞

for every test function ϕ. This follows easily from the following theorem.

Theorem 6.2.22. If F ∈ D (R) and F = 0, then F is a constant function. Proof: Let ϕ ∈ D(R). Using the notation of the proof of Theorem 6.2.21, we have  ∞ ϕ(x)dx, F, ϕ = F, Kϕ0 + ϕ1  = F, Kϕ0  + F, ϕ1  = F, ϕ0  −∞

because

$ % F, ϕ1  = − F , ψ = 0.

Thus, F is the regular distribution generated by the constant function C = F, ϕ0 .
In applications we often need to consider distributions on an open subset of RN .

(The space of test functions on an open subset of RN ) Let  be an open subset of RN . By the space of test functions on , we mean the

Definition 6.2.23.

6.2 Distributions

299

space of all smooth functions defined on  with support contained in some compact subset of . Elements of D() are called test functions on . Note that the function



(x ϕ(x) = e 0

2 −1)−1

if |x| < 1, otherwise,

is not a member of D ((−1, 1)). On the other hand, the function ψ(x) = ϕ(αx) is a member of D ((−1, 1)) for every α > 1.

(Convergence in D()) Let ϕ1 , ϕ2 , . . . and ϕ be test functions on . We say that the sequence {ϕn } converges to ϕ in D(), denoted by ϕn → ϕ in D(), if the following two conditions are satisfied:

Definition 6.2.24.

(a) ϕ1 , ϕ2 , . . . and ϕ vanish outside some compact set S ⊂ , (b) Dα ϕn → Dα ϕ uniformly on  for every multi-index α .

(The space of distributions on an open subset of RN ) By D () we denote the space of all continuous linear functions on D(). Members of D () are called distributions on .

Definition 6.2.25.

Definition 6.2.26.

(Multiplication of distributions by smooth functions) Let F ∈ D (), and let g be a smooth function on . By the product of g and F , we mean the distribution defined by gF, ϕ = F, gϕ.

It is easy to show that the product of a distribution and a smooth function is well defined.

Theorem 6.2.27. (Leibniz formula) Let F ∈ D (), and let g be a smooth function on . Then D (gF) = α

α  1

β≤α

β1



αN ... βN

where β ≤ α means βj ≤ αj for j = 1, . . . , N. We leave the proof as an exercise.

 Dβ gDα−β F,

300

Chapter 6

Partial Differential Equations

In the following sections, distributions are often used in a less precise fashion and ideas not mentioned in this section are used. The arguments can be made precise and formal definitions can be given, but it goes beyond the scope of this book.

6.3 Sobolev Spaces The study of partial differential equations naturally involves function spaces which are defined not only in terms of the functions themselves, but also of their derivatives. The Banach space of all bounded continuous functions on cl , where  is an open set in RN , with the norm of uniform convergence, is rather unsuitable for the study of partial differential equations. For example, if ∂ 2u ∂ 2u + · · · + =f 2 ∂x12 ∂xN with f continuous, it is generally not true that u ∈ C 2 (). Sobolev spaces turn out to be much more useful for the study of partial differential equations. In this section, we give the basic definitions and discuss some properties of Sobolev spaces. Several results are stated without proofs. An extensive treatment of Sobolev spaces can be found in Adams and Fournier (2003) or Ziemer (1989). Throughout this section,  stands for an open set in RN .

(Sobolev space W m,p ()) Let m ≥ 0 be an integer and 1 ≤ p ≤ ∞. The Sobolev space W m,p () is defined to be the set of all functions u ∈ Lp () such that Dα u ∈ Lp () for all |α| ≤ m.

Definition 6.3.1.

Since a function u ∈ Lp () need not be differentiable, the meaning of the condition Dα u ∈ Lp () requires an explanation. If u ∈ Lp (), then u can be identified with a regular distribution Gu ∈ D (). The distribution Gu has derivatives of all orders and Dα Gu ∈ D (). If Dα Gu is a regular distribution corresponding to a function in Lp (), then we simply write Dα u ∈ Lp (). It is easy to see that W m,p () is a vector space. The functional     α p 1/p   D u (6.20) u m,p, = |α|≤m 

is a norm in W m,p () if 1 ≤ p < ∞. For p = ∞ the norm is defined by   (6.21) u m,∞, = max Dα u∞ . 0≤|α|≤m

6.3 Sobolev Spaces

301

Example 6.3.2. For 1 ≤ p < ∞, we have  u 1,p, =

 1/p N     ∂u p   |u| + ,  ∂x  p



 u 2,p, =

k=1



k

  1/p N   N     ∂u p   ∂ 2 u p   +   |u| + .       ∂xk  ∂xk ∂xl p

k=1




k,l=1

Note that, if m1 > m2 , then W m1 ,p () ⊂ W m2 ,p () ⊂ Lp ()

(6.22)

u m1 ,p, ≥ u m2 ,p, ≥ u Lp ()

(6.23)

and for any u ∈ W

m1 ,p

().

Theorem 6.3.3. The Sobolev space W m,p () is a Banach space. If p < ∞, it is separable. The following property of Sobolev spaces is of fundamental importance for applications.

Theorem 6.3.4. The differential operator Dα is a continuous mapping from W m,p () to W m−|α|,p (), for any |α| ≤ m.

Proof: Let Dj denote the derivative with respect to the jth variable. Then, for any u ∈ W m,p (), we have      p Dα Dj upp ≤ Dβ upp = u p Dj u m−1,p, = m,p, , L () L () |α|≤m−1

|β|≤m

(6.24) which proves the theorem in the case of differential operators of order one. The general case follows easily.


Theorem 6.3.5. Let f be a function on  such that Dα f is bounded on  for every |α| ≤ m. Then multiplication by f is a continuous mapping from W m,p () into W m,p (). A mapping ϕ :  → RN is called a C r -diffeomorphism (r ∈ N), if the following conditions are satisfied: (a) Dα ϕ is continuous in  for all |α| ≤ r; (b) ϕ is one-to-one in ; (c) The Jacobian of ϕ does not vanish in .

302

Chapter 6

Partial Differential Equations

If ϕ is a C r -diffeomorphism in , then it is invertible and ϕ −1 is a C r ´ and Taylor (2001)). diffeomorphism in ϕ() (see, for example, Mikusinski

Theorem 6.3.6. Let  be bounded, m ∈ N, and let ϕ be a C m -diffeomorphism

in  defined in a neighborhood of cl . Then the mapping u → u ◦ ϕ −1 is a continuous linear operator from W m,p () into W m,p (ϕ()). The case when p = 2 is of special interest to us, because W m,2 () is an inner product space. The space W m,2 () is defined as a subspace of L2 (). However, it is not a closed subspace of L2 () and therefore, it is not a Hilbert space with respect to the inner product of L2 (). It turns out that W m,2 () is a Hilbert space if the inner product is defined by   Dα u(x)Dα v(x)dx. (6.25) u, vm, = |α|≤m 

The space W m,2 () is isomorphic to the space H m () (defined in Example 3.3.8) and the inner product (6.25) yields the norm (6.20) for p = 2. It is natural to consider the space Lp () as a special case of the Sobolev space with m = 0. We denote the Lp ()-norm of a function by · 0,p, . When  = RN , the space H m (RN ) can also be defined via the Fourier transform. If u ∈ H m (RN ), then Dα u ∈ L2 (RN ) for all |α| ≤ m. Thus, the Fourier transform of Dα u is well defined and given by 
F Dα u (ω) = i|α| ωα9 u(ω) u(ω)} ∈ L2 (RN ) for all |α| ≤ m. Conversely, if u ∈ L2 (RN ) and and so {ωα9 α {ω 9 u(ω)} ∈ L2 (RN ) for all |α| ≤ m, then u ∈ H m (RN ). We can express this in a more convenient form using the following lemma.

Lemma 6.3.7. For every m ∈ N, there exist positive constants K1 and K2 such that  m   α 2   ω  ≤ K2 1 + ω 2 m , K1 1 + ω 2 ≤

(6.26)

|α|≤m

for all ω ∈ RN .

Proof: We recall ω 2 = ω12 + ω22 + · · · + ωN2 and |ω|α = |ω1 |α1 · · · |ωN |αN . Using a simple induction argument on m, it follows that the same powers of  ω occur in (1 + ω 2 )m and |α|≤m |ωα |2 with different coefficients. Since the number of terms is finite and depends only on m, the inequalities (6.26) follow.
In view of this lemma, we can define the space H m (RN ) as follows:  
   m   H m RN = u ∈ L2 RN : 1 + ω 2 9 u(ω) ∈ L2 RN .

(6.27)

6.4 Fundamental Solutions

303

It follows from the Plancherel theorem that the norm · m,RN in H m (RN ) is equivalent to the norm 

2  m  u(ω) dω 1 + ω 2 9

u H m (RN ) =

1/2

.

(6.28)



The advantage of this definition is that (6.27) can be used to define H s (RN ) and the norm for all real numbers s ≥ 0. We return to the Sobolev spaces W m,p (). The map   ∂u ∂u u → u, ,..., (6.29) ∂x1 ∂xN is an isometry of W 1,p () into (Lp ())N+1 , where the latter space is equipped with the norm 1/p N+1  p ur 0,p, (6.30) u = r=1

for u = (ur ) ∈ (L ()) . If 1 ≤ p < ∞, the space of test functions D () is dense in Lp (). Moreover, if φ ∈ D (), then Dα φ ∈ D () for every α, and thus D () ⊂ W m,p () for any m and p. p

N+1

Theorem 6.3.8. D(RN ) is a dense subspace of W m,p (RN ) for any integer m ≥ 0

and 1 ≤ p < ∞.

If  is a proper subspace of RN , then this theorem need not hold. However, we have the following result.

Theorem 6.3.9. C ∞ () is a dense subspace of W m,p () for any integer m ≥ 0

and 1 ≤ p < ∞.

6.4 Fundamental Solutions and Green’s Functions for Partial Differential Equations Consider a partial differential operator L of order m in N variables  L= Aα D α ,

(6.31)

|α|≤m

where α = (α1 , . . . , αN ) is a multi-index, the αn ’s are non-negative integers, |α| = α1 +· · ·+αN , Aα = Aα1 ,α2 ,...,αN (x1 , x2 , . . . , xN ) are functions in RN (pos-

304

Chapter 6

Partial Differential Equations

sibly constant), and 

∂ D = ∂x1

α1

α



∂ ... ∂xN

αN

=

∂x1α1

∂ |α| αN . . . . ∂xN

The formal adjoint of L is L∗ v =



(−1)|α| Dα (Aα v).

(6.32)

|α|≤m

For example, the most general linear partial differential operator of order two in two independent variables is  L= Aα D α (6.33) |α|≤2

= A2,0 (x1 , x2 )

∂2 ∂2 ∂2 + A1,1 (x1 , x2 ) + A0,2 (x1 , x2 ) 2 2 ∂x1 ∂x2 ∂x1 ∂x2

+ A1,0 (x1 , x2 )

∂ ∂ + A0,1 (x1 , x2 ) + A0,0 (x1 , x2 ). ∂x1 ∂x2

(6.34)

When seeking a solution of the equation L(x) = Y , we may be interested in a solution which is a differentiable function, or just a function with Dα understood as generalized derivatives, or, finally, we may seek a solution which is a distribution. For this reason, the solutions of such an equation are classified as follows:

Definition 6.4.1. (a) (Classical solution). Let f be a function on RN . Every function u on RN which is sufficiently differentiable so that  Aα D α u |α|≤m

is well defined as a function and such that the equation  Lu = Aα D α u = f

(6.35)

|α|≤m

is satisfied is called a classical solution of (6.35). (b) (Weak solution). By a weak solution of (6.35), we mean a function on RN which need not be sufficiently differentiable to make Lu meaningful in the classical sense. In this case, f may be a function or a distribution.

6.4 Fundamental Solutions

305

(c) (Distributional solution). Let f ∈ D (RN ). Every u ∈ D (RN ) satisfying (6.35) is called a distributional solution of (6.35). Note that if f in (6.35) is a singular distribution, then the equation cannot have a classical solution. The remarkable fact is that including distributions one can generate new solutions of classical equations (equations where f is a function). Some classical equations may not even have a classical solution but can have distributional solutions.

Example 6.4.2. The differential equation xu = 0 has a weak solution of the form u(x) = c1 H(x) + c2 , where H is the Heaviside function and c1 and c2 are arbitrary constants. If c1 = 0, this is not a classical solution because H is not differentiable. However, u(x) = c2 is a classical solution.


Example 6.4.3. The differential equation x2 u = 0

(6.36)

has the same classical and weak solutions as xu = 0. Equation (6.36) has also distributional solutions u(x) = c1 δ(x) + c2 H(x) + c3 , where c1 , c2 , and c3 are arbitrary constants. Clearly, the distributional solutions have no analogues in the classical theory of differential equations.


Example 6.4.4. The nonhomogeneous equation u

(x) = δ

(x) has distributional solutions of the form u(x) = δ(x) + c1 x + c2 , where c1 and c2 are arbitrary constants.




It may be noted that, for a large class of problems, all generalized solutions of (6.35) turn out to be classical solutions. Equations of the form LG = δ

(6.37)

are of particular interest. Suppose G is a distribution satisfying (6.37). Then, for any distribution f with compact support, the convolution f ∗ G is well

306

Chapter 6

Partial Differential Equations

defined and L(f ∗ G) =

 |α|≤m

=



Aα Dα (f ∗ G)   Aα f ∗ D α G

|α|≤m

=f ∗



 Aα D G α

|α|≤m

= f ∗δ =f. Thus, if G is a solution of LG = δ, then f ∗ G is a solution of Lu = f . This explains the importance of the equations Lu = δ, at least in the context of existence of solutions of partial differential equations.

Definition 6.4.5.

(Fundamental solution) By a fundamental solution of a differential operator L, we mean a distributional solution of the equation Lu = δ . Clearly, a fundamental solution is not unique.

Example 6.4.6. It is easy to check that the Heaviside H function is a fundamental solution of the operator L = d/dx. If c is any constant, then H(x) + c is also a fundamental solution of L.
Consider the following boundary value problem Lu = f

in 

(6.38)

with one of the following boundary conditions: u = 0 on ∂,

(6.39)

∂u = 0 on ∂, ∂n

(6.40)

or ∂u + au = 0 on ∂, (6.41) ∂n where  is a bounded open set in RN and ∂ is the boundary of . The problem of finding the solution of (6.38) subject to the boundary condition (6.39) is called the homogeneous Dirichlet problem (Johann Peter Gustav Lejeune Dirichlet (1805–1859)). The problem of solving (6.38) satisfying the boundary condition (6.40) is known as the homogeneous Neumann problem. Finally, the

6.4 Fundamental Solutions

307

problem of finding the solution of (6.38) subject to the boundary condition (6.41) is called the Robin (or mixed) problem. A fundamental solution G of the differential operator L satisfies (6.37), but G need not satisfy the given boundary conditions. A fundamental solution which satisfies the homogeneous conditions is known as the Green’s function.

Definition 6.4.7. (Green’s function) By the Green’s function G of the operator L with homogeneous boundary conditions, we mean a fundamental solution of the equation LG = δ satisfying those boundary conditions. The remarkable fact is that the Green’s function can be used to solve the boundary value problems given by (6.38) to (6.41). The most important linear partial differential equations arise in problems of applied mathematics, mathematical physics, and engineering sciences. Included here are only examples of three-dimensional equations of most common interest:

Example 6.4.8. (The Laplace equation) ∇ 2 u = 0,

(6.42)

where ∇ 2u =

∂ 2u ∂ 2u ∂ 2u + 2 + 2. ∂x2 ∂y ∂z

The operator ∇ 2 is known as the Laplace operator (Pierre-Simon Laplace (1749–1827)). It is also denoted by . This equation is satisfied by the electrostatic potential in the absence of charges, by the gravitational potential in the absence of mass, by the equilibrium displacement of a membrane with a given displacement of its boundary, by the velocity potential for an inviscid, incompressible, irrotational homogeneous fluid in the absence of sources and sinks, by the temperature in steady-state heat flow in the absence of sources and sinks, and in many other situations.


Example 6.4.9. (The Poisson equation) ∇ 2 u = −f (x, y, z),

(6.43)

where f is a given function. The Poisson equation (Siméon Denis Poisson (1781–1840)) is satisfied by the electrostatic potential in the presence of charge, by the gravitational potential in the presence of distributed matter, by the equilibrium displacement of

308

Chapter 6

Partial Differential Equations

a membrane under distributed forces, by the velocity potential for an inviscid, incompressible, irrotational, homogeneous fluid in the presence of distributed sources or sinks, by the steady-state temperature in the presence of thermal sources or sinks, and in many other physical situations.


Example 6.4.10. (The nonhomogeneous wave equation) ∂ 2u − ∇ 2 u = −f (x, y, z). ∂t 2

(6.44)

This equation and the corresponding homogeneous form arise in a large number of physical situations. Some of these problems include the vibrating string, vibrating membrane, acoustic problems for the velocity potential for the fluid flow through which sound can be transmitted, longitudinal vibrations of an elastic rod or beam, and both electric and magnetic fields in the absence of charge and dielectric.


Example 6.4.11. (The nonhomogeneous heat (or diffusion) equation) ∂u − ∇ 2 u = f (x, y, z). ∂t

(6.45)


Example 6.4.12. (Telegrapher’s equation) ∂ϕ ∂ 2ϕ ∂ 2ϕ + a , + bϕ = ∂t 2 ∂t ∂x2

(6.46)

where a and b are constants. This equation arises in the study of propagation of electrical signals in a cable transmission line. Both the current I and the voltage V satisfy an equation of the form (6.46). This equation also arises in the propagation of pressure waves in the study of pulsatile blood flow in arteries and in the one-dimensional random motion of bugs along a hedge.


Example 6.4.13. (The inhomogeneous Helmholtz’s equation) (Hermann Ludwig Ferdinand von Helmholtz (1821–1894)) ∇ 2 ψ + λψ = −f (x, y, z),

(6.47)

where λ is a constant. This is essentially a time-independent inhomogeneous wave equation (6.44) with λ as the separation constant.


Example 6.4.14. (The biharmonic wave equation) ∇ 4ψ −

1 ∂ 2ψ = 0. c2 ∂t 2

(6.48)

6.4 Fundamental Solutions

309

In elasticity theory, the displacement of a thin elastic plate in small vibrations satisfies this equation. When ψ is independent of time t, (6.48) reduces to the so called biharmonic equation ∇ 4 ψ = 0.

(6.49)

This is the equilibrium equation for the distribution of stress in an elastic medium satisfied by Airy’s stress function ψ (George Biddell Airy (1801– 1892)). In fluid dynamics, the equation is satisfied by the stream function ψ in an incompressible viscous fluid flow.


Example 6.4.15. (The time-independent Schrödinger equation in quantum mechanics) (Erwin Rudolf Josef Alexander Schrödinger (1887–1961))
2 2 ∇ ψ + (E − V)ψ = 0, 2m

(6.50)

where m is the mass of the particle whose wave function is ψ, h = 2π
is the universal Planck’s constant, V is the potential energy, and E is a constant. If V = 0, (6.50) reduces to the Helmholtz’s equation.


Example 6.4.16. (The Klein–Gordon equation)
u + d2 u = 0,

(6.51)

where 1 ∂2 c2 ∂t 2 is the d’Alembertian operator and d is a constant.
= ∇2 −

(6.52)


The Laplacian has many remarkable features similar to a Sturm–Liouville operator in one spatial dimension. The following differential identity can be verified by direct computation: u∇ 2 v = ∇ · (u∇v) − (∇u) · (∇v). The integration over a domain  yields    2 u∇ vdτ = ∇ · (u∇u)dτ − (∇u) · (∇v)dτ. 



(6.53)

(6.54)



Application of the divergence theorem to the first integral on the right with the vector field (u∇v) and using the fact that ∇v · n is the directional derivative ∂v/∂n, we obtain the Green’s first identity    ∂v 2 u∇ vdτ = u ds − (∇u) · (∇v)dτ. (6.55)  ∂ ∂n 

310

Chapter 6

Partial Differential Equations

In particular, if u = v, (6.55) becomes    ∂u u∇ 2 udτ = u ds − |∇u|2 dτ.  ∂ ∂n 

(6.56)

Note that this is the higher dimensional analogue of the integration by parts formula  b  b uu

dx = [uu ]ba − (u )2 dx. (6.57) a

a

Interchanging u and v in (6.53) and subtracting from (6.53) gives u∇ 2 v − v∇ 2 u = ∇ · (u∇v − v∇u).

(6.58)

Integrating this equality over  and using the divergence theorem yields the Green’s second identity       2 ∂v ∂u 2 u −v u∇ v − v∇ u dτ = ds. (6.59) ∂n ∂n ∂ This formula can be also interpreted as the higher dimensional analogue of the integration by parts formula  b (uv

− vu

)dx = [uv − vu ]ba (6.60) a

or the integral form of the Lagrange identity as stated in Theorem 5.9.5. From (6.56) we note that if the boundary conditions on u are such that the integral over ∂ vanishes, then the operator −∇ 2 is positive definite. It also follows from (6.59) that ∇ 2 is a formally self-adjoint operator. Integrating once with respect to t, it is easy to check that the operator ∂/∂t has an adjoint −∂/∂t. Consequently, the adjoint of the operator (∇ 2 − ∂/∂t) is (∇ 2 + ∂/∂t), and (∇ 2 − ∂ 2 /∂t 2 ) is formally self-adjoint. Finally, the Green’s identities (6.56) and (6.59) can be generalized for operators more general than the Laplace operator by using the following identity:     v ∇ 2 u + a · ∇u + bu − u ∇ 2 v − ∇ · (av) + bv = ∇ · (v∇u − u∇v + auv), (6.61) where u and v are arbitrary differentiable functions, a is a vector field, and b is a constant scalar. This identity can be written in terms of differential operators L and L∗ as vLu − uL∗ v = ∇ · (v∇u − u∇v + auv). Application of the divergence theorem gives     ∗ vLu − uL v dτ = ∇ · (v∇u − u∇v + auv)dτ 



(6.62)

6.4 Fundamental Solutions

 =

311

(v∇u − u∇v + auv) · nds ∂

   ∂v ∂u − u + an uv ds, v = ∂n ∂n ∂

(6.63)

where ∂u/∂n = n · ∇u and an = a · n. This is an obvious extension of (6.59) and includes a special case when L is self-adjoint. It follows from the definition of L and L∗ that Lv = L∗ v only if a · ∇v = −∇ · (av) = −a · ∇v − v∇ · a. This gives ∇ · a = 0 when v = 1. If we assume v to take values x, y, and z successively, then ak = 0, k = 1, 2, 3. Thus, L is self-adjoint if and only if a = 0. We can also conclude that every self-adjoint operator in the sense of the definition (6.62) has the form Lu = ∇ 2 u + bu. Finally, it also follows from the vector identity

that

∇u · ∇v + u∇ 2 v = ∇ · (u∇v)

(6.64)

  v ∇ 2 u + bu + ∇u · ∇v − buv = ∇ · (v∇u).

(6.65)

Therefore, if L is self-adjoint, application of the divergence theorem gives    ∂u (6.66) vLudτ + (∇u · ∇v − buv)dτ = v ds.   ∂ ∂n This is clearly a generalization of (6.55). The second integral on the left-hand side of (6.66) expressed as a bilinear form  E(u, v) = (∇u · ∇v − buv)dτ (6.67) 

is called the Dirichlet integral of the operator L. If L is elliptic and b < 0, then E(u, v) is strictly positive since    E(u, u) = (∇u)2 − bu2 dτ > 0 (6.68) 

for any differentiable function u which does not vanish identically in . This property of the Dirichlet integral plays a fundamental role in the theory of elliptic partial differential equations. However, this feature is not universally applicable to all elliptic equations. For example, the inhomogeneous Helmholtz’s equation (6.47) with λ ≥ 0 has an associated Dirichlet integral, which is not positive.

312

Chapter 6

Partial Differential Equations

In general, if u and v are functions with continuous derivatives of order m, that is, Dα u and Dα v are continuous functions for every multi-index α with |α| ≤ m, then it can be shown that vLu − uL∗ v = div J(u, v),

(6.69)

where L and L∗ are given by (6.31) and (6.32), and J is a vector bilinear form in u and v involving only derivatives of u and v of order m − 1 or less. Equation (6.69) is an extension of the one-dimensional version of Green’s formula for partial differential operators. The integral form of (6.69) is the most general Green’s identity     n · Jds, (6.70) vLu − uL∗ v dτ = 

∂

where  is a bounded region in RN with sufficiently smooth boundary ∂, and n is the unit outward normal drawn to the surface element ds. We now illustrate the form of the Green’s second identity associated with several important differential operators. For the Helmholtz operator, we have L = L∗ = ∇ 2 + k2 , and the Green’s identity (6.63) takes the form       ∂v ∂u ∗ −u ds. v vLu − uL v dτ = ∂n ∂n  ∂

(6.71)

For the wave operator L = L∗ = c2

∂2 ∂2 − 2 2 ∂x ∂t

and the Green’s identity, (6.63) assumes the form    ,     ∂v ˆ ∂u ∂v ˆ ∂u ∗ 2 ˆ −u i− v −u j · nds. c v vLu − uL v dτ = ∂x ∂n ∂t ∂t  ∂ (6.72) Finally, for the diffusion operator L=K

∂ ∂2 − 2, ∂t ∂x

L∗ = −K

∂ ∂2 − 2, ∂t ∂x

and the Green’s identity, (6.63) takes the form   ,    ∂v ˆ ∂u ∗ ˆ ˆ i + Kuv j · nds. u −v vLu − uL v dτ = ∂x ∂n  ∂

(6.73)

6.4 Fundamental Solutions

313

Without a rigorous discussion, we shall illustrate the use of the Green’s function in finding the solution of the boundary value problems (6.38) to (6.41) for a linear inhomogeneous partial differential equation. Without loss of generality, we consider the problem in R3 . As before, the Green’s function G(x, ξ) of this problem satisfies the equation LG(x, ξ) = δ(x − ξ).

(6.74)

Physically, the Green’s function G(x, ξ) represents the effect at the point x of a Dirac delta function source at the point x = ξ. Multiplying (6.74) by f (ξ) and integrating over the volume V of the ξ space so that dv = dξ dηdζ , we find   L G(x, ξ)f (ξ)dv = δ(x − ξ)f (ξ)dv = f (x). (6.75) V

V

Comparing (6.38) and (6.75), we see that the solution of (6.38) can be written as  u(x) = G(x, ξ)f (ξ)dv. (6.76) V

Clearly, this is valid no matter how many components x may have. Accordingly, the Green’s function method can be applied, in principle, to any linear constant coefficient inhomogeneous partial differential equation in any number of independent variables. However, although a neat formulation has been developed, in practice, the construction of a Green’s function is not an easy problem. We next illustrate by the following examples the application of the Green’s function to linear constant coefficient inhomogeneous partial differential equations.

Example 6.4.17. The solution of the Poisson equation −∇ 2 u = f (x, y, z) is given by

(6.77)

 u(x, y, z) =

R3

G(x, ξ)f (ξ)dξ,

(6.78)

where the Green’s function G(x, ξ) of −∇ 2 is G(x, ξ) =

1 1 . 4π |x − ξ|

(6.79)

To find the fundamental solution for −∇ 2 , we need to solve the equation −∇ 2 G(x, ξ) = δ(x − ξ )δ(y − η)δ(z − ζ ),

(6.80)

314

Chapter 6

Partial Differential Equations

where x = ξ. We apply the three-dimensional Fourier transform with respect to x, y, and z  1 ˜ G(x, ξ)e−iκ·x dx (6.81) G(κ, ξ) = (2π )3/2 R3 to Equation (6.80) to obtain ˜= κ 2G

1 e−iκ·ξ (2π )3/2

(6.82)

where κ = (k, l, m) is the Fourier transform variable. The inverse Fourier transform gives the solution  dκ 1 eiκ·(x−ξ) 2 G(x, ξ) = 3 (2π ) R3 κ  dκ 1 = eiκ·r 2 , (2π )3 R3 κ

(6.83)

where r = |x − ξ|. Integral (6.83) can be evaluated using spherical coordinates in the κ variables. We choose the polar axis along the r direction and denote the spherical coordinates by κ, θ, ϕ. Then κ·r = κr cos θ where r = |r|. Then (6.83) becomes 1 G(x, ξ) = (2π )3



∞

 κ dκ



0



π

2

0

0

eiκr cos θ

dϕ κ2

∞

1 2 sin κr dκ 2 (2π ) 0 κr 1 1 = , = 4π r 4π |x − ξ| =

2π

sin θ dθ

(6.84)

provided r > 0. This completes the proof. In electrodynamics, the fundamental solution (6.84) has a well-known interpretation. It is essentially the potential at the point x produced by a unit point charge at the point ξ. This is what can be expected from a physical point of view because δ(x − ξ) is the charge-density corresponding to a unit point charge at ξ. The solution of (6.77) is   f (ξ, η, ζ ) 1 u(x, y, z) = G(x, ξ)f (ξ)dξ = dξ dηdζ. (6.85) 3 3 4π |x − ξ| R R The integrand in (6.85) consists of the given charge distribution f (x) at x = ξ and the Green’s function G(x, ξ). Physically, G(x, ξ)f (ξ) represents the

315

6.4 Fundamental Solutions

resulting potentials due to elementary point charges, and the total potential due to a given charge distribution f (x) is then obtained by the integral superposition of the resulting potentials. This is the so-called principle of superposition.


Example 6.4.18. The fundamental solution of the two-dimensional Helmholtz equation −∇ 2 G + λ2 G = δ(x − ξ )δ(y − η), is G(x, ξ) =

1 2π



∞

−∞ < x, y < ∞

rJ0 [r{(x − ξ )2 + (y − η)2 }1/2 ]dr . r 2 + λ2

0

(6.86)

(6.87)

It is convenient to introduce the change of variables x∗ = x − ξ and y∗ = y − η. Consequently, (6.86) assumes the form, dropping the asterisks, Gxx + Gyy − λ2 G = −δ(x)δ(y). Now we apply the double Fourier transform  1 ˜ e−iκ·x G(x, y)dx G(κ) = 2π R2

(6.88)

(6.89)

to (6.88) to obtain the solution 1 1 ˜ , G(κ) = 2 2π (κ + λ2 ) where κ = (k, l). The inverse Fourier transform gives the solution   −1 1 eiκ·x κ 2 + λ2 dκ. G(x, y) = 2 4π R2

(6.90)

(6.91)

In terms of the polar coordinates x = ρ cos θ, y = ρ sin θ, k = r cos ϕ, l = r sin ϕ, this integral has the representation G(x, y) =

1 4π 2

 0

∞

rdr r 2 + λ2



2π

eirρ cos(ϕ−θ ) dϕ.

0

The second integral can be expressed in terms of the Bessel function as 

2π 0

eirρ cos ϕ dϕ = 2π J0 (rρ).

316

Chapter 6

Partial Differential Equations

Hence, the solution becomes G(x, y, ξ, η) =

1 2π



∞ 0

rJ0 (rρ)dr . r 2 + λ2

(6.92)

Thus, in terms of the original coordinates, the fundamental solution is  ∞ rJ0 [r{(x − ξ )2 + (y − η)2 }1/2 ]dr 1 . G(x, ξ) = 2π 0 r 2 + λ2 Accordingly, the solution of the Helmholtz equation (∇ 2 − λ2 )u = −f (x, y) is

 u(x, y) =

R2

G(x, ξ)f (ξ)dξ,

(6.93)

where G(x, ξ) is given by (6.87). Since the integral in (6.92) does not exist for λ = 0, the Green’s function for the two-dimensional Poisson equation (6.86) cannot be obtained from (6.92). Instead, we differentiate (6.92) with respect to ρ to obtain  ∞ 2 ∂G 1 r J0 (rρ)dr = ∂ρ 2π 0 r 2 + λ2 which is, for λ = 0, 1 ∂G = ∂ρ 2π



∞ 0

J0 (rρ)dr = −

1 . 2πρ

It then follows that 1 ln ρ. 2π In terms of the original coordinates, we have G(ρ, θ ) = −

G(x, y, ξ, η) = −

 1  ln (x − ξ )2 + (y − η)2 . 4π

(6.94)

This is the Green’s function for the two-dimensional Poisson equation ∇ 2 = −f (x, y). Consequently, the solution of this equation is    1 f (ξ, η) ln (x − ξ )2 + (y − η)2 dξ dη. u(x, y) = − (6.95) 4π R2


Example 6.4.19. We find the fundamental solution of the Helmholtz equation 

   1 ∂ 2 ∂u r + k2 u = 0, ∇ +k u= 2 r ∂r ∂r 2

2

0 < r < ∞,

(6.96)

317

6.4 Fundamental Solutions

with the radiation condition lim r(ur + iku) = 0.

r→∞

(6.97)

In this case, the Green’s function must satisfy ∇ 2 G + k2 G =

δ(r) . 4π r2

(6.98)

Clearly, G satisfies 2 Grr + Gr + k2 G = 0 for r > 0 r or (rG)rr + k2 (rG) = 0. This equation admits a solution of the form rG = Aeikr + Be−ikr or eikr e−ikr +B . r r In order for G to satisfy the radiation condition, we need A = 0 and thus G=A

G=B

e−ikr . r

(6.99)

To determine B, we find     ∂G B −ikr 1 lim ds = − lim e + ik ds = 1, ε→0 S ∂n ε→0 S r r ε ε from which we obtain B = −1/(4π ). Consequently, the fundamental solution G(r) for (6.96) is e−ikr . (6.100) 4π r Physically, this represents outgoing spherical waves radiating away from the source at the origin. With a point source at a point ξ , the Green’s function G has the representation G(r) = −

G(x, ξ) = −

e−i|x−ξ|k , 4π |x − ξ|

where x and ξ are position vectors in R3 .

(6.101)

318

Chapter 6

Partial Differential Equations

Finally, when k = 0, these results reduce exactly to the known solution for the Poisson equation.


Example 6.4.20. The acoustic (or light) waves are incident from z = −∞ onto

a solid screen at z = 0 with a tiny hole at the origin. Describe the behavior of the waves after they pass through the hole. We seek solution of the wave equation utt = c2 ∇ 2 u

in the three-dimensional half-space z > 0. We assume that for z < 0, waves are propagating in from z = −∞, and hence, we have the representation u=

u0 −i(kz−ωt) , e ik

k = ω/c.

At the solid screen (z = 0), ∂u/∂z = 0, and at the hole, we take ∂u = u0 e−iωt . ∂z To determine u for z > 0, we seek solutions in the form u = (x)e−iωt so that  satisfies the Helmholtz equation  2  ∇ + k2  = 0 with the boundary conditions ∂ = 0 at the solid screen (z = 0), ∂z ∂ = u0 at the hole. ∂z Now the problem can be solved using Example 6.4.19.




Example 6.4.21. (Green’s function of the one-dimensional wave equation) We consider the one-dimensional inhomogeneous wave equation utt − c2 uxx = p(x, t),

x ∈ R, t > 0,

(6.102)

with the initial and boundary conditions u(x, 0) = 0,

ut (x, t) = 0 for x ∈ R,

u(x, t) → 0 as |x| → ∞.

(6.103) (6.104)

In this case, the Green’s function G must satisfy the equation Gtt − c2 Gxx = δ(x)δ(t),

x ∈ R, t > 0

(6.105)

6.4 Fundamental Solutions

319

and the initial and boundary conditions (6.103) and (6.104). We apply the joint Laplace and Fourier transform defined by  ∞  ∞ 1 −ikx : G(k, s) = √ e dx e−st G(x, t)dt. (6.106) 2π −∞ 0 The solution of the transformed problem is 1 1 : . G(k, s) = √ 2 2π (s + c2 k2 )

(6.107)

The inverse Laplace and Fourier transform gives the solution G(x, t) =

 1  H ct − |x| , 2c

(6.108)

where H is the Heaviside unit step function. With a source at (ξ, τ ), the Green’s function takes the form G(x, t; ξ, τ ) =

 1  H c(t − τ ), |x − ξ | . 2c

(6.109)

This function is also called the Riemann function for the wave equation. It follows from (6.109) that G = 0 unless the point (x, t) lies within the characteristic cone defined by the inequality c(t − τ ) > |x − ξ |. Finally, the solution of the inhomogeneous wave equation (6.102) is given by  t  ∞ u(x, t) = dτ G(x, t; ξ, τ )p(ξ, τ )dξ 0

=

1 2c

−∞





t

dτ 0

∞ −∞

  H c(t − τ ), |x − ξ | p(ξ, τ )dξ.

Since H = 1 for x − c(t − τ ) < ξ < x + c(t − τ ) and zero outside, the solution is  x+c(t−τ )  1 t u(x, t) = dτ p(ξ, τ )dξ. (6.110) 2c 0 x−c(t−τ )


Example 6.4.22. (Green’s function of the one-dimensional diffusion equation) We consider the inhomogeneous one-dimensional diffusion equation ut − Kuxx = q(x, t),

x ∈ R, t > 0

(6.111)

with the initial and boundary conditions u(x, 0) = 0 for x ∈ R

and

u(x, t) → 0 as |x| → ∞, t > 0.

(6.112)

320

Chapter 6

Partial Differential Equations

The Green’s function must satisfy the equation Gt − KGxx = δ(x)δ(t),

x ∈ R, t > 0,

(6.113)

with the initial and boundary conditions (6.112). Using the joint Laplace and Fourier transform defined by (6.106), we obtain the solution of the transformed problem as 1 1 : . (6.114) G(k, s) = √ 2π (s + Kk2 ) The joint inverse transform gives the solution   1 x2 G(x, t) = √ . exp − 4Kt 4π Kt

(6.115)

Note that G(x, t) is an even function of√x for all t > 0. The amplitude (or peak height) of√G decreases inversely with Kt, whereas the width of the peak increases with Kt. The spatial distribution of G(x, t) is√Gaussian, and, hence, it can easily be drawn against x for different values of 2 Kt. If the source is located at (ξ, τ ) instead of (0, 0), the corresponding Green’s function assumes the form   1 (x − ξ )2 G(x, t; ξ, τ ) = √ exp − . (6.116) 4K(t − τ ) 4Kπ(t − τ ) Thus, the solution of (6.111) is given by  t  ∞ dτ q(ξ, τ )G(x, t; ξ, τ )dξ. u(x, t) = 0

−∞

(6.117)

Finally, the solution of the inhomogeneous diffusion equation (6.111) with inhomogeneous initial data u(x, 0) = f (x),

x∈R

can be expressed as  t   ∞ f (ξ )G(x, t; ξ, 0)dξ + dτ u(x, t) = −∞

0

(6.118)

∞

−∞

q(ξ, τ )G(x, t; ξ, τ )dξ,

where G(x, t; ξ, τ ) is given by (6.116).

(6.119)


Example 6.4.23. (Green’s function of the one-dimensional Klein–Gordon equation) We consider the inhomogeneous Klein–Gordon equation utt − c2 uxx + d2 u = p(x, t),

x ∈ R, t > 0,

(6.120)

6.4 Fundamental Solutions

321

with the initial and boundary conditions u(x, 0) = 0 = ut (x, 0)

for x ∈ R,

u(x, t) → 0 as |x| → ∞, t > 0,

(6.121) (6.122)

where c and d are constants. The Green’s function G associated with this problem satisfies the equation Gtt − c2 Gxx + d2 G = δ(x)δ(t),

x ∈ R, t > 0,

(6.123)

with the same initial and boundary data (6.121) and (6.122). Application of the joint Laplace and Fourier transform yields the solution for the transformed problem as 1 1 : , G(k, s) = √ 2 2π (s + α 2 )

(6.124)

where α 2 = c2 k2 + d2 . The joint inverse transform leads to the solution  1 −1 sin αt G(x, t) = √ F α 2π

 ∞ sinct k2 + d2  1 c2

= eikx dk 2 2π c −∞ d 2 k + c2

,   d 22 1 2 c t − x H ct − |x| = J0 2c c  , d 22  1 2 c t −x if |x| < ct, J = 2c 0 c  0 if |x| > ct.

In the limit as d → 0, the Green’s function of the Klein–Gordon equation reduces to that of the standard wave equation (6.102). If the source is located at (ξ, τ ), the corresponding Green’s function takes the form  , d 2  1 c (t − τ )2 − (x − ξ )2 if |x − ξ | < c(t − τ ), J0 G(x, t; ξ, τ ) = 2c c  0 if |x − ξ | > c(t − τ ). (6.125) Finally, the solution of (6.120) with the initial data u(x, 0) = f (x)

and

ut (x, 0) = g(x),

x ∈ R,

(6.126)

322

Chapter 6

Partial Differential Equations

can be written as 



t

u(x, t) =

dτ 0

+



∞

−∞

p(ξ, τ )G(x, t; ξ, τ )dξ

∞

−∞

 g(ξ )G(x, t; ξ, 0) − f (ξ )Gτ (x, t; ξ, 0) dξ, (6.127)

where G = 0 for τ > t, and, hence, the integration with respect to τ extends only to t. It also follows from (6.125) that G is nonzero when |x − ξ | < c(t − τ ), which is equivalent to x −c(t −τ ) < ξ < x +c(t −τ ). Consequently, the double integral in (6.127) becomes 



t

dτ 0

∞ −∞



t

=

p(ξ, τ )G(x, t; ξ, τ )dξ 

x+c(t−τ )

dτ 0

p(ξ, τ )J0 x−c(t−τ )

, d 2 c (t − τ )2 − (x − ξ )2 dξ. c

(6.128)

Note that G(x, t; ξ, 0) is nonzero in (x − ct, x + ct) and vanishes outside of this interval. Consequently, 

∞ −∞

g(ξ )G(x, t; ξ, 0)dξ =

1 2c



x+ct

g(ξ )J0 x−ct

, d 2 2 c t − (x − ξ )2 dξ. (6.129) c

In terms of the Heaviside function, the Green’s function (6.125) can be written as ,   1 d 2 2 2 G(x, t; ξ, τ ) = J0 c (t − τ ) − (x − ξ ) H x + ct − (ξ + cτ ) 2c c   × H ξ + ct − (x + cτ ) . (6.130) This gives , td 2 2

d 2 2 2 2 c t − (x − ξ ) J0 c t − (x − ξ ) G(x, t; ξ, τ ) τ =0 = 2 c





× H(x − ξ + ct)H(ξ − x + ct) , 1 d 2 2 2 c t − (x − ξ ) + J0 2 c  × δ(x − ξ + ct)H(ξ − x + ct)  + δ(ξ − x + ct)H(x − ξ + ct) .

323

6.5 Elliptic Boundary Value Problems

In view of this result combined with the property of the delta function with H(2ct) = 1, we obtain  ∞ f (x)Gτ (x, t; ξ, 0)dξ − −∞

td =− 2



x+ct

J1

d c

c2 t 2 − (x − ξ )2



f (ξ )dξ c2 t 2 − (x − ξ )2 ,  1 ∞ d 22 2 J0 c t − (x − ξ ) + 2 −∞ c   × δ(x − ξ + ct)H(ξ − x + ct) + δ(ξ − x + ct)H(x − ξ + ct) f (ξ )dξ     1 td x+ct J1 dc c2 t 2 − (x − ξ )2 f (ξ )dξ + f (x − ct) + f (x + ct) . =− 2 x−ct 2 c2 t 2 − (x − ξ )2 (6.131) x−ct

Combining (6.128), (6.129), and (6.131), solution (6.127) takes the final form u(x, t) =

 1 f (x − ct) + f (x + ct) 2 ,  d 2 2 1 x+ct 2 J0 c t − (x − ξ ) g(ξ )dξ + 2c x−ct c    td x+ct J1 dc c2 t 2 − (x − ξ )2 f (ξ )dξ − 2 x−ct c2 t 2 − (x − ξ )2  x+c(t−τ ) ,  d 2 1 t dτ J0 c (t − τ )2 − (x − ξ )2 p(ξ, τ )dξ. + 2 0 c x−c(t−τ ) (6.132)

If d = 0, this solution reduces to that of the Cauchy problem for the inhomogeneous wave equation.


6.5 Weak Solutions of Elliptic Boundary Value Problems We consider the Dirichlet problem for the second-order elliptic operator −∇ 2 u = f

in ,

u = 0 on ∂,

(6.133)

where  ⊂ RN is a bounded open set, ∂ is the boundary of , and f ∈ C () is a given function.

324

Chapter 6

Partial Differential Equations

By definition, a classical solution u of this problem is a function u ∈ C 2 (cl ), which satisfies (6.133) at every point. We assume that u is a classical solution and multiply Equation (6.133) by ϕ ∈ D (), and then we integrate to obtain   − ϕ∇ 2 udτ = ϕfdτ. (6.134) 



Since ϕ = 0 on ∂, an application of the Green’s first identity to (6.134) yields   ∇u · ∇ϕdτ = f ϕdτ (6.135) 



for every ϕ ∈ D (). This does not require any information on the second derivatives of u. On the other hand, if f ∈ / C () the problem (6.133) does not have a classical solution. It is then necessary to generalize the solution in an appropriate manner. If f ∈ L2 (), Equation (6.135) makes sense if ∇u ∈ L2 (). If u ∈ H01 (), where H01 () is the subspace of H 1 () consisting of functions vanishing on ∂, and if the derivatives ∂u/∂xk are considered in the generalized sense, then it follows from the definition of the Sobolev space that ∂u/∂xk ∈ L2 (). Then if u ∈ H01 () and u satisfies (6.135), then it is a weak solution of (6.133). Since H01 () is the closure of D (), it is a dense subspace of H01 (). Therefore, solving Equation (6.135) is equivalent to finding u ∈ H01 () such that (6.136) ∇u, ∇ϕ = f , ϕ for all ϕ ∈ D (),  where ·, · is the inner product in L2 (): ϕ, ψ =  ϕψ. Equation (6.136) is known as the variational or weak formulation of the problem (6.133).

Theorem 6.5.1. Let  be a bounded open subset of RN , and let f ∈ L2 (). Then

there exists a unique weak solution u ∈ H01 () satisfying (6.136). Furthermore, u ∈ H01 () is a solution of (6.136) if and only if J(u) = min J(v),

(6.137)

v∈H01 ()

where 1 J(v) = 2



 ∇v · ∇vdτ − 

f vdτ.

(6.138)



Proof: In order to apply the Lax–Milgram theorem 4.3.16, we set H = H01 () and, for u, v ∈ H01 (),

 a(u, v) =

∇u · ∇vdτ. 

(6.139)

6.5 Elliptic Boundary Value Problems

325

We first show that a(·, ·) is coercive, that is, there exists a positive constant K such that ,   |∇u|2 dτ ≥ K u 21 = K |∇u|2 dτ + u2 dτ for all u ∈ H. 





This readily follows from Friedrichs’ first inequality   2 |∇u| dτ ≥ α u2 dτ, u ∈ H, 

(6.140)



where α is a positive constant. Thus,    1 1 2 2 |∇u| dτ = |∇u| dτ + |∇u|2 dτ 2 2      1 α ≥ |∇u|2 dτ + u2 dτ ≥ K u 21 , 2  2  where K = min{1/2, α/2} and u ∈ H. To prove the boundedness of a(·, ·), we note that     a(u, u) = |∇u|2 dτ ≤ |∇u|2 + u2 dτ = u 21 . 

(6.141)

(6.142)



Thus, a(·, ·) is bounded, symmetric, and coercive. So, by the Lax– Milgram theorem 4.3.16, there exists a unique weak solution of Equation (6.136). We next consider the Neumann boundary value problem −∇ 2 u + bu = f

in ,

(6.143a)

∂u =0 ∂n

on ∂,

(6.143b)

where  ⊂ RN is a bounded open set, n is the exterior unit normal to ∂, and b is a non-negative constant. According to Green’s first identity (6.55), if u is a classical solution, then u ∈ H01 () and it satisfies the equation    ∂u ∇u · ∇vdτ + v∇ 2 udτ = v ds.   ∂ ∂n Or, equivalently, by (6.143a) and (6.143b),    ∇u · ∇vdτ + buvdτ = f vdτ 

for every v ∈ H01 ().





(6.144)

326

Chapter 6

Partial Differential Equations

If f ∈ L2 (), then we define a weak solution of (6.143a) and (6.143b) as u ∈ H01 () satisfying (6.144). Consider the bilinear from associated with the operator A = −∇ 2 + b:    a(u, v) = (∇u · ∇v) + buv dτ. (6.145) 

Clearly, a is a bilinear form on H 1 () and  a(u, v) = (∇u · ∇v + buv)dτ 



≤ max(1, b1 )

(∇v · ∇v + uv)dτ 

= M(u, v) ≤ M u v , where 0 < b ≤ b1 , M = max(1, b1 ), and a is continuous. On the other hand,    ∇u · ∇u + bu2 dτ ≥ min(1, b0 ) u 2 , a(u, u) = 

where 0 < b0 ≤ b. Therefore, a is a continuous and coercive bilinear form. Then, by the Lax–Milgram theorem, there exists a unique solution u ∈ H 1 () such that a(u, v) = f , v

(6.146)

for all v ∈ H 1 (). This u is called the weak solution of the equation Au = f , that is, u is the unique solution of the Neumann boundary value problem (6.143a) and (6.143b). Furthermore, the solution minimizes the functional    1  2 J(v) = (6.147) ∇v · ∇v + bv dτ − f vdτ. 2  


Example 6.5.2. Consider the boundary value problem −∇ 2 u + a0 u = f

in  ⊂ R2 ,

(6.148a)

u=0

on ∂,

(6.148b)

where a0 is a positive constant. Set Tu = −∇ 2 u + a0 u. Define an inner product in H01 ()  u, v = (ux vx + uy vy + uv)dxdy, 

(6.149)

327

6.5 Elliptic Boundary Value Problems

a bilinear form in H01 ()



a(u, v) = v, Tu =

  v −∇ 2 u + a0 u dxdy,

(6.150)

f vdxdy.

(6.151)



and a functional on H01 ()

 I(v) = 

A quadratic form for this problem can be defined in H01 () by  ,   1
 2 1 2 2 I(u) = a(u, u) − I(u) = ux + uy + a0 u − fu dxdy. 2  2 The bilinear form a is symmetric, bounded, and positive definite. The boundedness follows from the Schwarz inequality + +       a(u, v) ≤ |ux |2 + |uy |2 dxdy |vx |2 + |vy |2 dxdy 

+



+ |u|2 dxdy

+ a0 

|v|2 dxdy 

≤ K u v , where K = max(1, a0 ). The positive definiteness follows from (6.150) by setting u = v:       2 2 a(u, u) = |∇u| + a0 u dxdy ≥ a0 |∇u|2 + |u|2 dxdy = α u 2 , 



where α = min(1, a0 ). Note that I(v) is bounded. Hence, it follows from the Lax–Milgram theorem that the problem a(u, v) = I(v) has a unique solution in H01 ().
We can generalize the preceding result to cover the case of second-order elliptic equations defined on an open bounded set  ⊂ RN with smooth boundary ∂. We now consider the boundary value problem Tu = f

in  ⊂ RN ,

(6.152a)

u=0

on ∂,

(6.152b)

where Tu = −

, N  ∂ ∂u aij + a0 u, ∂xi ∂xj

i,j=1

328

Chapter 6

Partial Differential Equations

aij ∈ C 1 (cl ), 1 ≤ i, j ≤ N, a0 ∈ C 1 (cl ), x = (x1 , . . . ,N ) ∈ RN . The differential operator T is said to be in divergence form. It is called uniformly elliptic if the ellipticity condition N 

  aij (x)ξi ξj ≥ K|ξ |2 = K ξ12 + · · · + ξN2

(6.153)

i,j=1

is satisfied for all ξ ∈ RN , x ∈ , and K is positive and independent of x and ξ . If f ∈ L2 (), a weak solution of (6.152a) and (6.152b) is given by   N  i,j=1

aij

∂u ∂v dτ + ∂xi ∂xj



 a0 uvdτ = 

f vdτ

(6.154)



for all v ∈ H01 (). It can readily be verified that every classical solution is a weak solution. Conversely, every sufficiently smooth weak solution is a classical solution. We next define a bilinear form in H01 () by a(u, v) =

  N

aij

 i,j=1

and the norm

∂u ∂v dτ + ∂xi ∂xj

 a0 uvdτ

(6.155)



  N    ∂u 2   u =   ∂x  dτ .

(6.156)

n

 n=1

If a0 (x) ≥ 0 for all x ∈ , then, in view of the ellipticity condition (6.153), a(u, u) =

  N  i,j=1

≥

  N  i,j=1

≥K

∂u ∂u aij dτ + ∂xi ∂xj aij

a0 u2 dτ 

∂u ∂u dτ ∂xi ∂xj

 N   ∂u 2 n=1



∂xn

= K u 2 .

It can be checked that the form a(u, v) is bounded in H01 (), that is,   a(u, v) ≤ M u v (6.157) for some constant M and all u, v ∈ H01 ().

6.6 Applications of the Fourier Transform

329

If a is symmetric, that is, aij = aji for all i, j ∈ N, then by the Lax–Milgram theorem, there exists a unique solution u ∈ H01 () such that a(u, v) = f , v

(6.158)

for all v ∈ H01 (). Consequently, u satisfies Equation (6.154). In other words, the unique solution u minimizes the functional 1 J(v) = 2

  N  i,j=1

∂v ∂v 1 aij dτ + ∂xi ∂xj 2



 a0 v dτ − 2



f vdτ

(6.159)



on H01 (). To define a weak solution through (6.154), it suffices to assume aij , a0 are bounded on . Hence, u is the weak solution of the equation Tu = f , that is, u is the unique weak solution of the elliptic boundary value problem (6.152a) and (6.152b). More generally, we consider the following second-order elliptic boundary value problem: Tu = f where

in  ⊂ RN ,

u = 0 on ∂,

(6.160)

, -  N N  ∂ ∂u ∂u ai + a0 u, aij + Tu = − ∂xi ∂xj ∂xi i,j=1

i=1

where the aij ’s satisfy the ellipticity condition (6.153) and ai ∈ C (cl ), 1 ≤ i ≤ N. A weak solution is a u ∈ H01 () satisfying a(u, v) = f , v

(6.161)

for every v ∈ H01 (), where a(u, v) =

  N  i,j=1

∂u ∂v aij dτ + ∂xi ∂xj

  N  i=1

∂u ai vdτ + ∂xi

 a0 uvdτ.

(6.162)



This bilinear form is not always symmetric. If it is symmetric, bounded, and coercive, then there exists a unique solution by the Lax–Milgram theorem.

6.6 Examples of Applications of the Fourier Transform to Partial Differential Equations Example 6.6.1. (One-dimensional diffusion equation with no sources or sinks) Consider the initial value problem for the one-dimensional diffusion

330

Chapter 6

Partial Differential Equations

equation with no sources or sinks: ut = Kuxx ,

−∞ < x < ∞, t > 0,

(6.163)

where K is a constant, with the initial data u(x, 0) = f (x).

(6.164)

This kind of problem can often be solved by the use of the Fourier transform  ∞ 1 e−ikx u(x, t)dx. u˜ (k, t) = √ 2π −∞ When the Fourier transform is applied to (6.163) and (6.164), we obtain u˜ t = −Kk2 u˜ , 1 u˜ (k, 0) = f˜ (k) = √ 2π



∞

−∞

e−ikx f (x)dx.

The solution of the transformed system is 2 u˜ (k, t) = f˜ (k)e−Kk t .

(6.165)

The inverse Fourier transform gives the solution  ∞ 1 2 f˜ (k)eikx−Kk t dk, u(x, t) = √ 2π −∞ which is, by the convolution Theorem 5.11.11,  ∞ 1 f (ξ )g(x − ξ )dξ, u(x, t) = √ 2π −∞ where

(6.166)

 ∞ 1 2 eikx−Ktk dk e =√ g(x) = F 2π −∞ ,    ∞ ix 2 x2 1 exp −Kt k − − =√ dk 2Kt 4Kt 2π −∞      1/2 1 x2 x2 π 1 =√ . exp − exp − =√ 4Kt 4Kt 2π Kt 2Kt −1




−Ktk2



Thus, the solution (6.166) becomes ,  ∞ 1 (x − ξ )2 u(x, t) = √ f (ξ ) exp − dξ. 4Kt 4π Kt −∞

(6.167)

6.6 Applications of the Fourier Transform

331

The integrand involved in the integral solution consists of the initial data f (x) and the Green’s function G(x, t): , 1 (x − ξ )2 G(x, t) = √ exp − . (6.168) 4Kt 4π Kt Since

, (x − ξ )2 = δ(x − ξ ), exp − lim √ t→0+ 4Kt 4π Kt if we let t → 0+, the solution becomes 1

(6.169)

u(x, 0) = f (x). Consider now the initial value problem ut = uxx + uyy ,

−∞ < x, y < ∞, t > 0,

u(x, y, 0) = f (x, y). The function

(6.170) (6.171)

, 2 1 x + y2 G(x, y, t) = exp − 4π t 4t

(6.172)

satisfies Equation (6.170). From this we can construct the formal solution ,  (x − ξ )2 + (y − η)2 1 dξ dη. (6.173) f (ξ, η) exp − u(x, y, t) = 4π t R2 4t Similarly, a formal solution of the initial value problem for the threedimensional diffusion equation ut = ∇ 2 u,

−∞ < x, y, z < ∞, t > 0,

u(x, y, z, 0) = f (x, y, z), is 1 u(x, y, z, t) = 8(π t)3/2

 R3

f (ξ, η, ζ )e−r

(6.174) (6.175)

2 /4t

dξ dηdζ,

(6.176)

where r2 = (x − ξ )2 + (y − η)2 + (z − ζ )2 .




Example 6.6.2. (One-dimensional wave equation) We obtain the d’Alembert solution of the Cauchy problem for a one-dimensional wave equation utt = c2 uxx ,

−∞ < x < ∞, t > 0,

u(x, 0) = f (x),

ut (x, 0) = g(x).

(6.177) (6.178)

332

Chapter 6

Partial Differential Equations

We apply the joint Fourier and Laplace transform defined by  ∞  ∞ 1 −ikx ˜u(k, s) = √ e dx e−st u(x, t)dt. 2π −∞ 0

(6.179)

The transformed Cauchy problem has the solution in the form u˜ (k, s) =

sf˜ (k) + g˜ (k) . s2 + c2 k2

(6.180)

The joint inverse transformation gives the solution 1 u(x, t) = √ 2π



 ˜ sf (k) + g˜ (k) e L dk, s2 + c2 k2 −∞ ∞

ikx

−1

(6.181)

where L−1 is the inverse Laplace transform operator. Finally, we obtain ,  ∞ 1 g˜ (k) sin ckt dk u(x, t) = √ eikx f˜ (k) cos ckt + ck 2π −∞  ∞  1 1 ikx  ickt =√ e e + e−ickt f˜ (k)dk 2π −∞ 2  ∞  1 1 1 ikx  ickt +√ e e − e−ickt g˜ (k)dk 2π 2ic −∞ k  x+ct   1 1 1 ∞ = f (x + ct) + f (x − ct) + √ g˜ (k)dk eikζ dζ 2 2c 2π −∞ x−ct  x+ct  1 1 = f (x + ct) + f (x − ct) + g(ζ )dζ. (6.182) 2 2c x−ct This is the classical d’Alembert solution (Jean Le Rond d’Alembert (1717– 1783)). It can be shown, by direct substitution, that it is the unique solution of the wave equation provided f is twice continuously differentiable and g is once differentiable. This essentially proves the existence and uniqueness of the d’Alembert solution. It can also be shown, by direct substitution, that the solution (6.182) is uniquely determined by the initial data. It is important to point out that the solution u depends only on the initial values at points between x − ct and x + ct and not at all on initial values outside this interval on the line t = 0. This interval is called the domain of dependence of the variables (x, t). Moreover, it can be proved that the solution depends continuously on the initial data, that is, the problem is well posed. In particular, if g(x) = 0, the d’Alembert solution (6.182) reduces to u(x, t) =

 1 f (x + ct) + f (x − ct) . 2

(6.183)

333

6.6 Applications of the Fourier Transform

Figure 6.3 Splitting of initial data with equal waves. Physically, this solution shows that the initial data u(x, 0) = f (x) are split into two identical traveling waves of half amplitude moving in opposite directions with constant velocity c as shown in Figure 6.3. In general, the physical significance of the d’Alembert solution can be investigated by rewriting the solution (6.182) in the form 1 1 u(x, t) = f (x − ct) − 2 2c



x−ct 0

1 1 g(ζ )dζ + f (x + ct) + 2 2c



x+ct

g(ζ )dζ 0

= (x − ct) + (x + ct),

(6.184)

where 1 1 (ξ ) = f (ξ ) − 2 2c



ξ

g(ζ )dζ, 0

1 1 (η) = f (η) + 2 2c



η

g(ζ )dζ. 0

Physically, (x − ct) represents a wave traveling in the positive x-direction with constant speed c without change of shape. Similarly, (x + ct) also represents a wave propagating in the negative x-direction with the same speed c without change of shape. Finally, the solution of the inhomogeneous wave equation (6.102) with the initial data (6.178) can be obtained by adding (6.110) to the d’Alembert solution (6.182) in the form   1 x+ct 1 u(x, t) = f (x + ct) + f (x − ct) + g(ζ )dζ 2 2c x−ct  x+c(t−τ )  1 t + dτ p(ζ, τ )dζ. (6.185) 2c 0 x−c(t−τ )


334

Chapter 6

Partial Differential Equations

Example 6.6.3. (Laplace’s equation in a half-plane) We consider the Dirichlet problem consisting of the Laplace equation uxx + uyy = 0,

−∞ < x < ∞, y ≥ 0

(6.186)

with the boundary conditions u(x, 0) = f (x), u(x, y) → 0 as r = We introduce the Fourier transform



1 u˜ (k, y) = √ 2π



∞ −∞

(6.187) x2 + y2 → ∞.

(6.188)

e−ikx u(x, y)dx

so that from (6.186) to (6.188), we find d2 u˜ − k2 u˜ = 0, dy2 u˜ (k, 0) = f˜ (k),

u˜ (k, y) → 0 as y → ∞.

Thus, the solution of the transformed system is u˜ (k, y) = f˜ (k)e−|k|y .

(6.189)

From the convolution Theorem 5.11.11, we obtain the solution  ∞ 1 f (ξ )g(x − ξ )dξ, u(x, y) = √ 2π −∞ where g(x) = F

−1




e

−|k|y



. =

y 2 . 2 π x + y2

Consequently, the solution (6.190) becomes  f (ξ )dξ y ∞ , u(x, y) = π −∞ (x − ξ )2 + y2

y > 0.

(6.190)

(6.191)

(6.192)

This is the well-known Poisson integral formula in the half plane. Note that ,  ∞  ∞ 1 y f (ξ ) lim+ f (ξ )δ(x − ξ )dξ, dξ = lim+ u(x, y) = y→0 y→0 π (x − ξ )2 + y2 −∞ −∞ where Cauchy’s definition of the delta function is used, that is, δ(x − ξ ) = lim+ y→0

1 y . π (x − ξ )2 + y2

335

6.6 Applications of the Fourier Transform

Figure 6.4 Isothermal curves representing a family of circular curves. This may be recognized as a solution of the Laplace equation for a dipole source at (x, y) = (ξ, 0). In particular, when   f (x) = T0 H a − |x| , the solution (6.192) reduces to  dξ yT0 a u(x, y) = π −a (ξ − x)2 + y2   ,  T0 −1 x + a −1 x − a = tan − tan π y y   2ay T0 tan−1 2 . = π x + y2 − a2

(6.193)

(6.194)

The curves in the upper half-plane for which the steady-state temperature is constant are known as isothermal curves. In this case, these curves represent a family of circular curves x2 + y2 − αy = a2

(6.195)

with centers on the y-axis and fixed end points on the x-axis at x = ±a as shown in Figure 6.4. Similarly, we can solve the Dirichlet problem for the three-dimensional Laplace equation in the half-space: uxx + uyy + uzz = 0,

−∞ < x, y < ∞, z ≥ 0,

(6.196)

with the initial and boundary conditions u(x, y, 0) = f (x, y),

(6.197)

336

Chapter 6

Partial Differential Equations

u(x, y, z) → 0 as r =



x2 + y2 + z2 → ∞.

Application of the double Fourier transform gives the solution  z f (ξ, η)dξ dη . u(x, y, z) = 2 2π R2 [(x − ξ ) + (y − η)2 + z2 ]3/2

(6.198)

(6.199)


Example 6.6.4. We find the potential for the irrotational two-dimensional inviscid fluid flow filling the half-space. For the irrotational motion, the curl of the velocity vector v is zero everywhere, so that v can be expressed in terms of the velocity potential ϕ by v = −∇ϕ. The continuity equation, div u = 0, reduces to the Laplace equation for ϕ: ∇ 2 ϕ = ϕxx + ϕyy = 0,

−∞ < x < ∞, y ≥ 0.

(6.200)

The fluid is introduced normally to the half-space through the strip |x| ≤ a of the plane y = 0. This gives the boundary conditions U f (x) 2 (u, v) = −(ϕx , ϕy ) → (0, 0) ϕy = −

on y = 0,

(6.201)

as y → ∞,

(6.202)

where U is a constant and f is a given function. Application of the Fourier transform  ∞ 1 e−ikx ϕ(x, y)dx ϕ(k, ˜ y) = √ 2π −∞

(6.203)

enables us to solve the problem. The formal solution has the form U ϕ(x, y) = √ 2π



∞ −∞

f˜ (k) ikx−|k|y e dk. |k|

(6.204)

√ In particular, if f (x) = H(a − |x|), then f˜ (k) = (2/ 2π )(sin ka)/k so that the solution becomes  ∞ sin ka ikx−|k|y U e ϕ(x, y) = dk. (6.205) 2π −∞ k|k| Thus, the velocity component in the y direction is given by  ∞ U sin ka ikx−|k|y dk v = −ϕy = Re e 2π k −∞

6.6 Applications of the Fourier Transform

=

U 2π

=

U 4π

 

337

∞

1 sin ka cos kxe−|k|y dk −∞ k ∞
−∞

sin k(x + a) − sin k(x − a)

 e−|k|y k

dk,

(6.206)

where Re stands for the real part. Using the result 

∞

sin αk 0

π y e−ky dk = − tan−1 , k 2 α

Solution (6.206) for v becomes , U y y −1 −1 tan − tan . v= 2π x−a x+a Similarly, for the x-component of the velocity, we obtain  r2 iU ∞ sin ka ikx−|k|y U e ln , dk = u = −ϕx = − 2π −∞ k 2π r1

(6.207)

(6.208)

where r12 = (x − a)2 + y2 and r22 = (x + a)2 + y2 . Introducing a complex potential w = ϕ + iψ, we obtain ∂ϕ dw ∂ϕ = −i = −u + iv, dz ∂x ∂y which can be written, by (6.206) to (6.208), in the form , U r1 z−a U dw = ln , ln + i(θ1 − θ2 ) = dz 2π r2 2π z + a

(6.209)

(6.210)

where tan θ1 = y/(x − a) and tan θ2 = y/(x + a). Integrating (6.210) with respect to z gives the complex potential w=

 U 2a + (z − a) ln(z − a) − (z + a) ln(z + a) . 2π

(6.211)


Example 6.6.5. (The Navier–Stokes equation) The Navier–Stokes equation (Claude Louis Marie Henri Navier (1785–1836), George Gabriel Stokes (1819– 1903)) in a viscous fluid of constant density ρ and constant kinematic viscosity ν with no external forces is Du 1 = − ∇p + ν∇ 2 u, Dt ρ

(6.212)

338

Chapter 6

Partial Differential Equations

where u = (u, v, w) is the local Eulerian fluid velocity at a point x = (x, y, z) and at time t, p(x, t) is the pressure, and the total derivative following the motion is D ∂ = + u · ∇, (6.213) Dt ∂t which consists of an unsteady term and a convective term. We next introduce the vorticity vector ω = (ξ, η, ζ ) in rectangular Cartesian coordinates ξ = wy − vz ,

(6.214a)

η = uz − wx ,

(6.214b)

ζ = vx − uy .

(6.214c)

1 u × curl u = ∇(u · u) − u · ∇u 2

(6.215)

Using the vector identity

with q2 = u · u, Equation (6.212) assumes the form   p 1 2 ∂u + u · ∇u = −∇ + q + ν∇ 2 u. ∂t ρ 2

(6.216)

Taking the curl of both sides of this equation, the pressure term disappears and hence we get ∂ω = curl(u × ω) + ν∇ 2 ω ∂t = −u · ∇ω + ω · ∇u + ν∇ 2 ω

(6.217) (6.218)

in which the continuity equations, ∇ · u = 0 and ∇ · ω = 0, are used. Equation (6.218) can be also written in the form Dω = ω · ∇u + ν∇ 2 ω. Dt

(6.219)

This equation (or its equivalent form (6.218)) is called the vorticity transport equation and represents the rate of change in vorticity ω, which is described by three terms on the right-hand side of (6.218). The first term, u · ∇ω, is the familiar rate of change due to convection of fluid in which the vorticity is nonuniform past a given point. The second term, ω · ∇u, describes the stretching of vortex lines, and the last term, ν∇ 2 ω, represents the rate of change of ω due to molecular diffusion of vorticity in exactly the way that ν∇ 2 u represents the contribution to the acceleration from the diffusion of velocity (or momentum).

6.6 Applications of the Fourier Transform

339

In the case of two-dimensional flow, ω is everywhere normal to the plane of flow, and ω · ∇u = 0. Equation (6.219) then reduces to the scalar equation Dω = ν∇ 2 ω, Dt

(6.220)

so that only convection and viscous conduction occur. In terms of the stream function ψ, where u = ψy , v = −ψx (and ω = −∇ 2 ψ) satisfy the continuity equation identically, Equation (6.218) assumes the form   ∂ψ ∂ ∂ψ ∂ ∂ + − (6.221) ∇ 2 ψ = ν∇ 4 ψ. ∂t ∂y ∂x ∂x ∂y In the steady-state ∂/∂t = 0, and if the velocity of the fluid is very small and the viscosity is very large, all terms on the left-hand side of (6.221) can be neglected in the first approximation. Consequently, (6.221) reduces to the biharmonic equation ∇ 4 ψ = 0.

(6.222)

We solve this biharmonic equation for the viscous fluid bounded by the plane y = 0 with the fluid introduced through a strip |x| < a of this plane with a given velocity. Thus, the boundary conditions are ψy = 0 on y = 0,   ψx = f (x)H a − |x| .

(6.223a) (6.223b)

Furthermore, we consider only solutions which tend to zero at large distance from the strip. Application of the Fourier transform to ψ(x, y) with respect to x reduces (6.222) and (6.223a) to (6.223b) to 2  2 d 2 ˜ y) = 0, (6.224) − k ψ(k, dy2 ψ˜ y = 0,

y = 0,

(6.225a)

ikψ˜ = f˜ (k),

y = 0,

(6.225b)

where 1 f˜ (k) = √ 2π



a −a

e−ikx f (x)dx.

Thus, the solution of the system (6.224) and (6.225a)–(6.225b) is ˜ y) = − ψ(k,

 i 1 + |k|y f˜ (k)e−|k|y . k

(6.226)

340

Chapter 6

Partial Differential Equations

By the means of the inverse Fourier transform combined with the convolution theorem, we obtain  ∞ 1 f (x )g(x − x )dx , (6.227) ψ(x, y) = √ 2π −∞ where

 ∞,  −|k|y+ikx i 1 − 1 + |k|y e dk g(x) = √ 2π −∞ k . , xy 2 x tan−1 + 2 = . π y x + y2

Finally, the solution of the boundary value problem is   ,

 (x − x )y 1 ∞ −1 x − x + tan f (x )dx . ψ(x, y) = π −∞ y (x − x )2 + y2 In particular, if f (x) = δ(x), the solution (6.229) becomes   , 1 xy −1 x . ψ(x, y) = tan + 2 π y x + y2

(6.228)

(6.229)




Example 6.6.6. In wave propagation problems in applied mathematics, a typical initial boundary problem is utt = −Lu

in ,

(6.230)

u=0

on ∂,

(6.231)

u = f (x, y)

at t = 0,

(6.232a)

ut = g(x, y)

at t = 0,

(6.232b)

where  is a bounded domain in R2 with a smooth boundary ∂, and L is the differential operator , 1 ∂u ∂u ∂ ∂ Lu = − p1 (x, y) − p2 (x, y) + q(x, y)u , (6.233) r(x, y) ∂x ∂x ∂y ∂y r, p1 , p2 , and q are continuous positive functions on the closure of , and p1 and p2 have continuous derivatives. We solve the problem in the Hilbert space L2 (), which consists of functions f (x, y) with the square of the norm   2 2 f = r(x, y)f (x, y) dxdy < ∞. (6.234) 

6.6 Applications of the Fourier Transform

341

We seek a separable solution of (6.230) in the form u(x, y, t) =

∞ 

an (t)ϕn (x, y),

(6.235)

n=1

where ϕn are the eigenfunctions of L with eigenvalues λn , so that Lϕn = λn ϕn . We also assume that f (x, y) =

∞ 

fn ϕn (x, y),

(6.236a)

gn ϕn (x, y).

(6.236b)

n=1

g(x, y) =

∞  n=1

A simple substitution of (6.235) into (6.230) gives ∞ 

a

n (t)ϕn (x, y) = −

n=1

∞ 

λn an (t)ϕn (x, y).

n=1

Thus, the functions an satisfy the equation a

n (t) + λn an (t) = 0 so that

(6.237)

√ sin λn t an (t) = An √ + Bn cos λn t. λn

To satisfy the initial conditions, we set An = gn and Bn = fn , and the final solution becomes √ ∞ ,  sin λn t u(x, y, t) = (6.238) fn cos λn t + gn √ ϕn (x, y). λn n=1 This clearly satisfies the given equation, as well as boundary and initial conditions.


Example 6.6.7. In a general diffusion process, a typical initial boundary problem is ut = −Lu

in  ⊂ R2 ,

(6.239)

u=0

on ∂,

(6.240)

u(x, y, t) = f (x, y)

at t = 0,

(6.241)

where L is given by (6.233).

342

Chapter 6

Partial Differential Equations

Following the argument similar to that employed in Example 6.6.6, we write the solution in the form u=

∞ 

an (t)ϕn (x, y),

n=1

and we assume that f (x, y) =

∞ 

fn ϕn (x, y).

n=1

Consequently, the functions an satisfy the equation a n (t) = −λn an (t) so that an (t) = An e−λn t . We choose An = fn to obtain the solution u(x, y, t) =

∞ 

fn e−λn t ϕn (x, y).

(6.242)

n=1

A simple check reveals that (6.242) satisfies the differential equation and the boundary and initial conditions.


Example 6.6.8. (The linearized Korteweg–de Vries equation) The linearized Korteweg–de Vries (KdV) (Diederik Johannes Korteweg (1848–1941)) equation for the free surface evolution η(x, t) in an inviscid water of constant depth h is ch2 ηxxx = 0, −∞ < x < ∞, t > 0, ηt + cηx + (6.243) 6 where c = gh is the shallow water speed. We solve Equation (6.243) with the initial condition η(x, 0) = f (x),

−∞ < x < ∞.

(6.244)

Application of the Fourier transform, with respect to x, to the KdV system gives the solution in the form ,  2 2 kh ˆ −1 , E(k, t) = f (k) exp ikct 6 where E(k, t) = F {η(x, t)}. The inverse Fourier transform gives ,   2  ∞ 1 cth ˆ k2 dk. η(x, t) = √ f (k) exp ik (x − ct) + 6 2π −∞

(6.245)

343

6.7 Exercises

In particular, if f (x) = δ(x), then (6.245) reduces to the Airy integral  2  ∞ , cth 1 cos k(x − ct) + (6.246) η(x, t) = k3 dk, 2π 0 6 which is, in terms of the Airy function,  2 −1/3 , 2 −1/3 cth cth Ai (x − ct) , η(x, t) = 6 6 where the Airy function Ai(z) is defined by    1 ∞ 1 3 Ai(z) = cos kz + k dk. π 0 3

(6.247)

(6.248)


6.7 Exercises 1. Let f and g be continuous functions on RN . Show that if   f (x)ϕ(x)dx = g(x)ϕ(x)dx RN

RN

for every ϕ ∈ C ∞ (RN ) with compact support, then f (x) = g(x) for every x ∈ RN . 2. Show that a test function ψ is of the form ψ(x) = (xϕ(x)) , where ϕ is a test function if and only if  ∞  0 ψ(x)dx = 0 and ψ(x)dx = 0. −∞

0

3. Show that D is a vector space. 4. Show that if ϕ, ψ ∈ D , then (a) f ϕ ∈ D for every smooth function f . (b) {ϕ(Ax)} ∈ D for every affine transformation A of RN onto RN . (c) ϕ ∗ ψ ∈ D . 5. Construct a test function ϕ such that ϕ(x) = 1 for |x| ≤ 1, and ϕ(x) = 0 for |x| ≥ 2. 6. Which of the following expressions define a distribution ?

344

Chapter 6

m

(a) f , ϕ =

n=1 ϕ

(b) f , ϕ = (c) f , ϕ =

(n)

m

(0).

n=1 ϕ(xn ), x1 , . . . , xm

∞

n=1 ϕ

(d) f , ϕ = (e) f , ϕ =

Partial Differential Equations

(n)

∞

∈ R are fixed.

(0).

n=1 ϕ(xn ), x1 , x2 , . . . ∈ R

m

n=1 ϕ

(n)

are fixed.

(xn ), x1 , . . . , xm ∈ R are fixed.

(f) f , ϕ = (ϕ(0))2 . (g) f , ϕ = sup ϕ. ∞ (h) f , ϕ = −∞ |ϕ(t)|dt. b (i) f , ϕ = a ϕ(t)dt.  (j) f , ϕ = ∞ n=1 ϕ(xn ), where limn→∞ xn = 0. D

D

7. Let ϕn −→ ϕ and ψn −→ ψ. Prove the following: D

(a) aϕn + bψn −→ aϕ + bψ for any scalars a, b. D

(b) f ϕn −→ f ϕ for any smooth function f defined on RN . D

(c) ϕn ◦ A −→ ϕ ◦ A for any affine transformation A of RN onto RN . D

(d) Dα ϕn −→ Dα ϕ for any multi-index α. 8. Is the convergence in D metrizable ? 9. Let f be a locally integrable function on RN . Prove that the functional F on D defined by  F, ϕ =

RN

fϕ

is a distribution. 10. Find the nth distributional derivative of f (x) = |x|. 11. Let fn (x) = sin nx. Show that fn → 0 in the distributional sense. 12. Let {fn } be the sequence of functions on R defined by  fn (x) =

0, n, 0,

if x < −1/2n; if −1/2n ≤ x ≤ 1/2n; if x > 1/2n.

Show that the sequence converges to the Dirac delta distribution.

6.7 Exercises

345

13. Show that the sequence of Gaussian functions on R defined by . n −n2 x2 e fn (x) = , n = 1, 2, . . . , π converges to the Dirac delta distribution. 14. Show that the sequence of functions on R defined by fn (x) =

sin nx , πx

n = 1, 2, . . . ,

converges to the Dirac delta distribution.

∞ 15. Let ϕ0 ∈ D (R) be a fixed test function such that −∞ ϕ0 (x)dx = 1. Show that every test function ϕ ∈ D (R) can be represented in the form ϕ = Kϕ0 + ϕ1 , where K is a constant and ϕ1 is a test function such that Moreover, the representation is unique.

∞

−∞ ϕ1 (x)dx

= 0.

16. Let  be an open set in RN . Show that if F ∈ D () and g is a smooth function on , then     ∂ ∂ ∂ (gF) = g F+g F , ∂xj ∂xj ∂xj for j = 1, . . . , N. 17. Prove the Leibniz formula for the product of smooth function and a distribution (Theorem 6.2.27). 18. Consider the two-dimensional diffusion equation ut − K∇ 2 u = f (x, y)δ(t),

(x, y) ∈ R2 , t > 0

with the initial and boundary conditions u(x, y, 0) = 0

for all (x, y) ∈ R2 , u(x, y, t) → 0 as r = x2 + y2 → ∞.

Show that the Green’s function is   G (x, y), (ξ, η) =

, 1 (x − ξ )2 + (y − η)2 exp − . 4π Kt 4Kt

Generalize the result for the n-dimensional diffusion equation.

346

Chapter 6

Partial Differential Equations

19. Consider the three-dimensional inhomogeneous wave equation utt − c2 ∇ 2 u = f (x, y, z, t),

(x, y, z) ∈ R3 , t > 0

with the initial and boundary conditions u(x, y, z, 0) = 0 = ut (x, y, z, 0), u(x, y, z, t) → 0 as x2 + y2 + z2 → ∞. Show that the Green’s function of this problem is   1 2 1 2 2 δ t− x +y +z . G(x, y, z, t) = c 4π c2 x2 + y2 + z2 20. Find the Green’s function G(x, y, z, t) for the three-dimensional Klein– Gordon equation where G(x, y, z, t) satisfies the equation Gtt − c2 ∇ 2 G + d2 G = δ(x)δ(y)δ(z)δ(t),

(x, y, z) ∈ R3 , t > 0

with the initial and boundary conditions G(x, y, z, 0) = 0 = Gt (x, y, z, 0) for all (x, y, z) ∈ R3 and G(x, y, z, t) → 0 as



x2 + y2 + z2 → ∞, t > 0.

21. Use the joint Laplace and Fourier transform to solve the inhomogeneous Cauchy problem utt − c2 uxx = p(x, t), u(x, 0) = f (x)

and

x ∈ R, t > 0, ut (x, 0) = g(x)

for all x ∈ R.

22. Apply the joint Fourier and Laplace transform to obtain the Green’s function for the wave equation Gtt − c2 Gxx = δ(x)δ(t),

−∞ < x < ∞, t > 0,

G(x, 0) = Gt (x, 0) = 0. 23. (a) Show that the fundamental solution G(x, ξ, t) for the Cauchy problem Gtt = c2 Gxx , G(x, 0) = 0,

−∞ < x < ∞, t > 0, Gt (x, 0) = δ(x − ξ ),

6.7 Exercises

347

is G(x, ξ, t) =

 1 H(x − ξ + ct) − H(x − ξ − ct) . 2c

(b) Use this fundamental solution to solve a more general wave problem utt = c2 uxx ,

−∞ < x < ∞, t > 0,

u(x, 0) = 0,

ut (x, 0) = g(x).

24. Prove the existence of the weak solution of the Dirichlet boundary value problem −∇ 2 u + cu = f

in  ⊂ R2 ,

u = 0 on ∂,

where c is a positive function of x and y. Show that the weak solution is given by     2 v −∇ u + cu dτ = f vdτ, 

where u, v



∈ H01 ().

25. Show that the Dirichlet problem for the biharmonic operator ∇ 4u = f u=

in ,

∂u =0 ∂n

f ∈ L2 (),

on ∂,

where  ⊂ RN , has a weak solution u ∈ H02 () given by   uvdτ = f vdτ for every v ∈ H02 (). 



26. Show that the boundary value problem −u + u = f

in RN ,

  f ∈ L2 RN ,

u → 0 as |x| → ∞ has a unique solution u ∈ H 1 (RN ) such that    ∇u · ∇vdτ + uvdτ = RN

for all v ∈ H 1 (RN ).

RN

RN

f vdτ

348

Chapter 6

Partial Differential Equations

27. Let  ⊂ RN be a bounded open set. Consider the Robin boundary value problem −u + u = f

in ,

f ∈ L2 (),

∂u + αu = 0 ∂n

on ∂,

α > 0.

Show that there exists a unique solution u ∈ H01 () such that a(u, v) = f , v for every v ∈ H01 (), where







∇u · ∇vdτ +

a(u, v) = 

uvdτ + α 

uvdτ

and

∂

u, v ∈ H01 ().

28. Use the Fourier transform method to show that the solution of the telegrapher’s problem utt + aut + bu = c2 uxx , u(0, t) = f (t),

−∞ < x, t < ∞,

ux (0, t) = g(t)

is u(x, t) =

1 2π



∞, ∞


 g˜ (k)
 f˜ (k) cos xα(k) + sin xα(k) eikx dk, α(k)

where α(k) =

b + ika − k2 , c2

and f˜ and g˜ are the Fourier transforms of f and g, respectively. 29. Find the solution of the Neumann problem in the half-plane uxx + uyy = 0,

−∞ < x < ∞, y > 0,

uy (x, 0) = g(x),

−∞ < x < ∞,

u is bounded as y → ∞,

u, ux vanish as |x| → ∞.

30. Find the solution of the system uxx + uyy = 0, u(x, 0) = f (x),

−∞ < x < ∞, 0 ≤ y ≤ a, u(x, a) = g(x).

6.7 Exercises

349

31. Solve the boundary value problem uxx + uyy = 0,

−∞ < x < ∞, 0 ≤ y ≤ a,

u(x, 0) = f (x),

uy (x, a) = 0.

32. Show that the solution of the slow motion of viscous fluid through a slit governed by the biharmonic Equation (6.222) with the boundary conditions   ψy = g(x)H a − |x| , ψx = 0 on y = 0 is y2 ψ(x, y) = π



∞ −∞

g(x )dx . (x − x )2 + y2

33. Use the analysis of Example 6.6.4 to find the solution of the twodimensional steady flow of an inviscid liquid through a slit in a plane rigid boundary y = 0. The problem is to find the velocity potential ϕ(x, y) satisfying the Laplace equation with the boundary conditions   ϕ = H a − |x| ,

  1/2  v = −ϕy = a2 − x2 H a − |x| on y = 0.

34. If E(u, v) is a bilinear form defined by the Dirichlet integral (6.67) of a self-adjoint operator L, prove the following: (a) E(αu + βv, αu + βv) = α 2 E(u, v) + 2αβE(u, v) + β 2 E(v, v), where α and β are constants. (b) (E(u, v))2 ≤ E(u, u)E(v, v), if b ≤ 0 and L is an elliptic operator. 35. (a) Show that the solution of the one-dimensional Schrödinger equation for a free particle of mass m i
is ψ(x, t) =

∂ψ
2 ∂ 2 ψ =− ∂t 2m ∂x2

  x2 N exp − 2 , b 2b

  i
t 1/2 , b = a2 + m

where N and b are constants. (b) Show that the Gaussian probability density is  2 x |N|2 exp − 2 |ψ| = ac c 2

350

Chapter 6

Partial Differential Equations

and its mean width is c δ= √ , 2

 
2 t 2 1/2 2 . where c = a + 2 2 ma

36. (a) Show that the solution of the one-dimensional Schrödinger Equation (6.50) with the finite square well potential, V(x) = V0 H(|x| − a), is (i) ψ1 (x) = A exp(κx), x ≤ −a, where κ 2 =

2m
2 (V0

− E) > 0,

(ii) ψ2 (x) = B sin kx + C cos kx, |x| < a, where k2 =

2mE
,

(iii) ψ3 (x) = D exp(−κx), x ≥ a, where A, B, C, and D are arbitrary constants. (b) Use the matching conditions at x = ±a to show that the eigenvalue equations are k cot ak = −κ and k tan ak = κ, (c) Discuss the even and odd solutions of the Schrödinger equation. 37. If gt (x) =

√1 2kt

2

x exp(− 4kt ), show that gt (x) ∗ gs (x) = gt+s (x).

Mathematical Foundations of Quantum Mechanics

CHAPTER

7

“The tool which serves as intermediary between theory and practice, between thought and observation, is mathematics; it is mathematics which builds the linking bridges and gives the ever more reliable forms. From this it has come about that our entire contemporary culture, in as much as it is based on the intellectual penetration and the exploitation of nature, has its foundations in mathematics. Already Galileo said: one can understand nature only when one has learned the language is mathematics and these signs are mathematical figures. . . . Without mathematics, the astronomy and physics of today would be impossible; these sciences, in their theoretical branches, virtually dissolve into mathematics.” David Hilbert “The research worker, in his efforts to express the fundamental laws of Nature in mathematical form, should strive mainly for mathematical beauty. He should take simplicity into consideration in a subordinate way to beauty. . . . It often happens that the requirements of simplicity and beauty are the same, but where they clash the latter must take precedence.” Paul Dirac

7.1 Introduction This chapter is devoted to the Hilbert space formalism of quantum mechanics in general terms without aiming at mathematical rigor or generality. The theory of Hilbert spaces and the so-called Hermitian operators are essential for the mathematical foundation of quantum mechanics. Although physical concepts and their interpretations underlying quantum mechanics are radi-

351

352

Chapter 7

Quantum Mechanics

cally different from those of classical physics, the relation between these two subjects is of special importance. Our aim is to discuss the basic concepts and postulates of quantum mechanics in terms of the results of the theory of operators in a Hilbert space.

7.2 Basic Concepts and Equations of Classical Mechanics In classical mechanics, matter is usually made up of point particles. We assume that a macroscopic system is made up of a number of electrically neutral particles, of which the ith particle has mass mi and coordinates (xi , yi , zi ). They move under the action of mutually interacting forces according to Newton’s laws of motion (Sir Isaac Newton (1643–1727)). With the forces F = (Fxi , Fyi , Fzi ) acting on the ith particle, Newton’s Second Law of motion has the form d2 (7.1) mi 2 (xi , yi , zi ) = (Fxi , Fyi , Fzi ). dt In vector notation, they are written as mi

d 2 ri = Fi , dt 2

(7.2)

where ri = (xi , yi , zi ) and Fi = (Fxi , Fyi , Fzi ). In general, the particles are assumed to be nonrelativistic, that is, their speed is much less than that of light. The total force acting on the particles is assumed to be irrotational (∇ × F = 0) and hence, can be derived from a potential energy function, V = V(r, t) using Fi = −∇i V,

(7.3)

where ∇i = (∂/∂xi , ∂/∂yi , ∂/∂zi ) and the potential V is a scalar function of all coordinates (three for each particle) and time t. A field of force is often called conservative, if it is irrotational and independent of time t. When a particle is under the action of conservative forces, it can be shown that the total work done by such forces in carrying a system about a closed path is zero. In view of (7.3), Newton’s equations (7.2) can be written in a more general form dpi , (7.4) −∇i V = Fi = dt where pi = mvi = m˙ri = (m˙xi , m˙yi , m˙zi ) is the linear momentum of the particle.

7.2 Basic Concepts and Equations

353

The main problem of classical dynamics is to describe the general motion of a particle (or a system of particles) when its initial position and velocity are prescribed and the total force acting on the particle (or system) is given. The kinetic energy Ti of the ith particle is a function of three timederivatives of position or velocities r˙ ≡ dr/dt and is defined by      1 dri dri 1  2 2 2 Ti = mi x˙ i + y˙ i + z˙i = mi · . 2 2 dt dt

(7.5)

Hence, the kinetic energy of the system of particles is T=

 i

Ti =

1 i

2

 mi

dri dt

   dri · . dt

(7.6)

Using (7.5), Newton’s Equations (7.4) can be expressed as a set of equations in terms of kinetic and potential energies for each particle:   d ∂T + ∇i V = 0. dt ∂ r˙i

(7.7)

Or, more explicitly, ,

     - , d ∂T d ∂T ∂V ∂V ∂V d ∂T , , + = 0. , , dt ∂ x˙ i dt ∂ y˙ i dt ∂ z˙i ∂xi ∂yi ∂zi

(7.8)

The total energy E of a particle is defined as the sum of its kinetic and potential energies and, in general, is a function of r, r˙ , and t, that is, E = T + V(r, t).

(7.9)

A dynamical function is called a constant of motion if it does not change in time as the particle moves along its trajectory. Such a function is said to be conserved. When a particle is under the action of conservative forces, its energy E is always a constant of the motion, that is, energy is conserved. From (7.9) and the assumption that V is independent of t, we obtain   dE dv dp = mv · + v · ∇V = v · + ∇V = 0, (7.10) dt dt dt by (7.4). Hence E is a constant of motion. The Lagrangian function L of the system is defined as the difference of the kinetic and potential energies. More precisely, L = T − V = L(xi , yi , zi , x˙ i , y˙ i , z˙i ),

(7.11)

354

Chapter 7

Quantum Mechanics

where the set of six arguments of L stands for one set of six arguments for each particle. Consequently, ∂L ∂T = ∂ x˙ i ∂ x˙ i

(7.12a)

∂V ∂L =− ∂xi ∂xi

(7.12b)

and

so that Equation (7.8) can be expressed in terms of the Lagrangian function L as   d ∂L ∂L = 0, (7.13a) − dt ∂ x˙ i ∂xi   d ∂L ∂L = 0, (7.13b) − dt ∂ y˙ i ∂yi   d ∂L ∂L = 0. (7.13c) − dt ∂ z˙i ∂zi These are clearly the transformed version of Newton’s equations and are known as the Lagrange equations of motion for each particle. There is no reason to restrict Newton’s equations to Cartesian coordinates. In fact, many dynamical problems that cannot be solved in Cartesian coordinates can be solved using other coordinate systems. So it would be more general and more convenient to develop a formalism equivalent to Newton’s (or Lagrange’s) equations which is independent of any particular system of coordinates. Thus, the transformation of Newton’s equations leads to an important idea of coordinate transformation which can be introduced by defining the Cartesian coordinates (xi , yi , zi ) in terms of new coordinates qj and time t where j = 1, 2, 3, . . . , 3n for n particles in a system. Since the configuration of the system at time t is specified by values of qj and t, xi = xi (q1 , q2 , . . . , q3n , t),

(7.14a)

yi = yi (q1 , q2 , . . . , q3n , t),

(7.14b)

zi = zi (q1 , q2 , . . . , q3n , t),

(7.14c)

where qj are called the generalized coordinates, which are independent so that there can be no functional relation between them. The only requirement is that the generalized coordinates be independent and that they uniquely determine the position of a particle. Conversely, a set of generalized coordinates may be defined in terms of the Cartesian variables by a transformation of the form qj = qj (r, t) or simply qj = qj (r). There can be no integrable linear relation connecting their increments δqj . The name generalized is justified because the

355

7.2 Basic Concepts and Equations

nature of the coordinates qj are not specified. They may be distances, angles, areas, or any other set of numbers such that when their values are known, the configuration of the system is uniquely prescribed. A dynamical system is called holonomic if there are no nonintegrable relations between δqj ’s. Otherwise, it is nonholonomic. Thus, in a holonomic system, not only are the coordinates qj independent, but so also are the infinitesimals δqj . In such a system, it is therefore possible to effect a displacement such that qk becomes qk + δqk while the remaining ones q1 , q2 , . . . , qk−1 , qk+1 , . . . , q3n retain constant values. It can be shown that Lagrange’s equations (7.13a) to (7.13c) have the same form in terms of qj for a holonomic system. To prove this result, we first compute x˙ i , y˙ i , z˙i from (7.14a) to (7.14c) to obtain x˙ i =

∂xi ∂xi ∂xi q˙ 1 + · · · + q˙ 3n + , ∂q1 ∂q3n ∂t

y˙ i =

∂yi ∂yi ∂yi , q˙ 1 + · · · + q˙ 3n + ∂q1 ∂q3n ∂t

z˙i =

∂zi ∂zi ∂zi . q˙ 1 + · · · + q˙ 3n + ∂q1 ∂q3n ∂t

Evidently, ∂ x˙ i ∂xi = , ∂ q˙ j ∂qj

(7.15a)

∂yi ∂ y˙ i = , ∂ q˙ j ∂qj

(7.15b)

∂zi ∂ z˙i = . ∂ q˙ j ∂qj

(7.15c)

The kinetic energy of the system is  1  2 mi x˙ i + y˙ i2 + z˙i2 2  1 = a11 q˙ 21 + a22 q˙ 22 + · · · + 2a12 q˙ 1 q˙ 2 + · · · + 2a1 q˙ 1 + · · · + a 2

T=

1  ajk q˙ j q˙ k , = 2 3n

3n

(7.16)

j=1 k=1

where ajk =



 mi

∂yi ∂yi ∂zi ∂zi ∂xi ∂xi + + ∂qj ∂qk ∂qj ∂qk ∂qj ∂qk

 = akj ,

(7.17a)

356

Chapter 7

Quantum Mechanics

 ∂xi ∂xi ∂yi ∂yi ∂zi ∂zi + + , aj = mi ∂qj ∂t ∂qj ∂t ∂qj ∂t  , ∂xi 2  ∂yi 2  ∂zi 2 + + . a= mi ∂t ∂t ∂t 



(7.17b) (7.17c)

It also follows from (7.15a) that x¨ i

    d ∂xi ∂ x˙ i ∂ x˙ i d ∂xi = x¨ i = − x˙ i x˙ i ∂qj ∂ q˙ j dt ∂ q˙ j dt ∂qj ,    ∂ 1 2 d ∂ 1 2 x˙ i x˙ i . − = dt ∂ q˙ j 2 ∂qj 2

Thus, it turns out that     ∂xi ∂yi ∂zi ∂T  d ∂T = mi x¨ i + y¨i + z¨ − dt ∂ q˙ j ∂qj ∂qj ∂qj ∂qj    ∂xi ∂yi ∂zi + Yi + Zi , = Xi ∂qj ∂qj ∂qj

(7.18)

where Xi , Yi , Zi are the components of the total force (internal and external) acting on the ith particle. In view of the fact that the system is holonomic, a displacement can be made such that qj becomes qj + δqj , while all other q’s remain constant. If Qj δqj is the work done in this displacement, then  Qj δqj = (Xi δxi + Yi δyi + Zi δzi )    ∂xi ∂yi ∂zi = Xi + Yi + Zi δqj . ∂qj ∂qj ∂qj Substituting this result into (7.18), we obtain   ∂T d ∂T = Qj . − dt ∂ q˙ j ∂qj

(7.19)

These are called Lagrange’s equations of motion for a holonomic dynamical system. A great advantage of these equations for the solution of dynamical problems is that forces which do no work, such as reactions at smooth hinges, do not appear. However, these are to be included in the usual equations of motion. An important modification of (7.19) can be made for conservative dynamical systems. In such cases, ∂V Qj = − , (7.20) ∂qj

7.2 Basic Concepts and Equations

357

where V is the potential energy of the system and is a function of the generalized coordinates q1 , q2 , . . . , q3n and possibly of the time t. Since V does not involve the velocities q˙ j , ∂V/∂ q˙ j = 0. Thus, Lagrange’s equation (7.19) can be written in terms of L as   d ∂L ∂L = 0, j = 1, 2, . . . , 3n. (7.21) − dt ∂ q˙ j ∂qj If the kinetic energy of the system T is a homogeneous quadratic function of velocities q˙ j , then 2T =

 ∂T  ∂L = q˙ j . q˙ j ∂ q˙ j ∂ q˙ j

(7.22)

Assuming that V does not involve the time t explicitly, it follows from (7.21) that , d  ∂L  d d ∂L ∂L (T + V) = (2T − L) = q˙ j − q˙ j − q¨ j dt dt dt ∂ q˙ j ∂ q˙ j ∂qj   ,  d ∂L ∂L − q˙ j = 0. (7.23) = q˙ j dt ∂ q˙ j ∂qj This shows that T + V is constant for a conservative system. To discuss the so-called Hamilton equations of motion (Sir William Rowan Hamilton (1805–1865)), we introduce the concepts of generalized momentum pj and generalized force Fj by pj =

∂L ∂ q˙ j

(7.24a)

Fj =

∂L . ∂qj

(7.24b)

and

Consequently, the Lagrange equations (7.21) become ∂L d = pj = p˙ j , ∂qj dt

(7.25)

where pj and qj are usually called conjugate variables. In Cartesian coordinates, this result reduces to the familiar equation for pxj : pxj =

∂L ∂T = = m˙xj . ∂ x˙ ∂ x˙

(7.26)

Any equations that will hold in any coordinate system are of the most interest and useful for applications. To develop such equations of motion, we now

358

Chapter 7

Quantum Mechanics

introduce a new function, called the Hamiltonian (or Hamilton’s function) H, which is defined in terms of the Lagrangian L and two conjugate variables pj and qj by  pj q˙ j − L, (7.27) H(p, q) = j

where L = L(qj , q˙ j , t) is, in general, a function of qj , q˙ j , t, and q˙ j enters through the kinetic energy as a quadratic term. Equation (7.24a) will give pj as a linear function of q˙ j . This system of linear equations involving pj and q˙ j can be solved to determine q˙ j in terms of pj , and then the q˙ j ’s can in principle be eliminated from (7.27). This essentially means that H can always be expressed as a function of pj , qj , and t so that H ≡ H(pj , qj , t). Thus, dH =

 ∂H ∂pj

dpj +

 ∂H ∂qj

dqj +

∂H dt. ∂t

(7.28)

On the other hand, differentiation of H in (7.27) with respect to t gives  d  ∂L d  ∂L d dH  d ∂L q˙ j pj − = pj q˙ j + qj − q˙ j − , (7.29) dt dt dt ∂qj dt ∂ q˙ j dt ∂t or dH =



pj d˙qj +



q˙ j dpj −

 ∂L  ∂L ∂L dqj − d˙qj − dt. ∂qj ∂ q˙ j ∂t

(7.30)

In view of (7.24a), this equation becomes dH =



q˙ j dpj −

 ∂L ∂L dqj − dt. ∂qj ∂t

(7.31)

Evidently, the two expressions of dH given in (7.28) and (7.31) must be equal so that the coefficients of the corresponding differentials can be equated to obtain ∂H , ∂pj

(7.32a)

∂L ∂H = , ∂qj ∂qj

(7.32b)

∂L ∂H = . ∂t ∂t

(7.32c)

q˙ j = −

−

Using the Lagrange equations (7.21) and (7.24a), the first two of the preceding equations become, for each j, dqj ∂H = , dt ∂pj

(7.33a)

7.2 Basic Concepts and Equations

dpj ∂H =− , dt ∂qj

j = 1, 2, . . . , 3n.

359 (7.33b)

These are known as Hamilton’s canonical equations of motion. They constitute a set of 6n coupled first-order equations that reflects the symmetry except for a negative sign. Thus, the Hamilton equations are completely equivalent to Lagrange’s Equations (7.13a) to (7.13c), which represent a set of 3n coupled second-order differential equations. These equations possess a unique solution if the initial data are prescribed at some time t = t0 . In other words, Hamilton’s equations completely determine the position and momentum at all times provided the initial data are given. This shows that the fundamental laws of classical mechanics are completely deterministic. The Lagrange–Hamilton theory can be employed to deduce (a) the law of conservation of energy and (b) that H is equal to the total energy. To derive (a), we assume that L, and therefore, H in (7.27) do not involve the time t explicitly. Consequently,  < dH d ; = pj q˙ j − L qj , q˙ j dt dt   ∂L ∂L q˙ j − q¨ j = pj q¨ j + p˙ j q˙ j − ∂qj ∂ q˙ j   = pj q¨ j + p˙ j q˙ j − p˙ j q˙ j − pj q¨ j = 0. This shows that H is constant. To prove the second property, we assume that the coordinate transformations (7.14a) to (7.14c) do not depend explicitly on time t. We note that T in L = T − V is given by (7.16) where the coefficients ajk are symmetric functions of the generalized coordinates qj . On the other hand, V is, in general, independent of q˙ j and hence, ∂L ∂T  = = ajk q˙ k . ∂ q˙ j ∂ q˙ j 3n

pj =

k=1

Thus, the Hamiltonian H becomes H=

3n  j=1

pj q˙ j − L =

3n  j=1

 q˙ j

3n 

 ajk q˙ k − L

j=1

= 2T − L = T + V. Thus, H is equal to the total energy. Since H was proved to be a constant, the sum of the kinetic and potential energies is constant. This is the celebrated law of conservation of energy.

360

Chapter 7

Quantum Mechanics

We consider a conservative holonomic system that is described by the generalized coordinates q1 , q2 , . . . , qn . For any complete set of specified initial conditions, each of the coordinates qj must be a single-valued function of time t. Thus, we assume that these functions are known and have the form q1 = q1 (t), q2 = q2 (t), . . . , qn = qn (t). Then these equations may be regarded as the parametric equations of a path in n-dimensional Euclidean space, and the motion of the system can be related to that of a point which moves along this path. Given the initial state, the motion in subsequent times is uniquely determined by Newton’s laws of motion. Therefore, there exists a unique path in the n-dimensional Euclidean space for a given set of initial data. It is of interest to compare this path with another one in the n-dimensional space, the two paths having the same end points and as such that they are traversed in the same time τ . Assuming that at any instant of time the difference between the positions of two points which trace out the two paths is infinitesimally small, we denote the variation between the two paths at any instant by δ. Both paths have the same end points so that δq1 (0) = δq1 (τ ) = · · · = δqn (0) = δqn (τ ) = 0. However, the variations δq1 (t), δq2 (t), . . . , δqn (t) do not vanish at any time t between 0 and τ . Then it follows that , ∂L ∂L δ q˙ j + δqj δL = ∂ q˙ j ∂qj   , ∂L d d ∂L δqj = (δqj ) + ∂ q˙ j dt dt ∂ q˙ j   d  ∂L δqj . = dt ∂ q˙ j Thus,



τ

δ 0

   d  ∂L Ldt = δLdt = δqj dt ∂ q˙ j 0 0 dt   τ  ∂L d δqj = ∂ q˙ j 0 , , ∂L ∂L δqj − δqj = 0, = ∂ q˙ j ∂ q˙ j t=τ t=0 



τ

τ

since both the paths have the same end points. Evidently,  τ Ldt = 0. δ

(7.34)

0

This result is well known as Hamilton’s variational principle provided the Lagrangian L = L(qj , q˙ j ). This principle was obtained from Newton’s law of motion.

7.2 Basic Concepts and Equations

361

Conversely, it can be shown that Newton’s laws can be derived from Hamilton’s principle if L = L(qj , q˙ j ). Clearly, , ∂L ∂L δL = δ q˙ j + δqj . ∂ q˙ j ∂qj By Hamilton’s principle   τ Ldt = 0=δ

   τ  ∂L ∂L δLdt = δ q˙ j + δqj dt ∂ q˙ j ∂qj 0 0 0 , -τ   τ , d  ∂L  ∂L ∂L = δqj dt + δqj , − δqj + dt ∂ q˙ j ∂qj ∂ q˙ j 0 0 τ

where the last result is obtained by integration by parts. Since δqj = 0 at t = 0 and t = τ , the last term vanishes and the preceding expression gives    τ , ∂L d ∂L − δqj dt = 0 ∂qj dt ∂ q˙ j 0 for all δqj and all τ . Thus, the integrand must vanish, which yields Lagrange’s equation   ∂L d ∂L = 0, (7.35) − dt ∂ q˙ j ∂qj which is the transformed version of Newton’s equations. Hence, Newton’s laws can be derived from these equations. Hamilton’s principle shows that motion according to Newton’s laws is distinguished  from all other kinds of motion by having the property that the integral Ldt for any given time interval has a stationary value. Hence, it is regarded as a fundamental principle of classical mechanics from which everything else can be derived.

Poisson’s Brackets in Mechanics The equations of motion for any canonical function F(pi , qi , t) can be expressed, using Hamilton’s Equations (7.33a) and (7.33b), as  3n  dF  ∂F ∂F ∂F q˙ i + = p˙ i + dt ∂qi ∂pi ∂t i=1

 3n   ∂F ∂H ∂F ∂H ∂F = − + ∂qi ∂pi ∂pi ∂qi ∂t i=1

362

Chapter 7

Quantum Mechanics

= {F, H} +

∂F , ∂t

(7.36)

where {F, H} is called the Poisson bracket of two functions F and H. If the canonical function F does not explicitly depend on time t, then ∂F/∂t = 0 so that (7.36) becomes dF = {F, H}. dt

(7.37)

In addition, if {F, H} = 0, then F is a constant of the motion. In fact, (7.37) really includes the Hamilton equations, which can be verified by setting F = pi , F = qi , or F = H. It readily follows the definition of the Poisson bracket that {qi , pj } = δij ,

(7.38a)

{qi , qj } = {pi , pj } = 0,

(7.38b)

where δij is the Kronecker delta notation. These are the fundamental Poisson brackets for the canonically conjugate variables pi and qi . Any relation involving Poisson’s brackets must be invariant under a canonical transformation. This is often used as an alternative definition of a canonical transformation. It can also be verified that the components of the angular momentum L = r × p satisfy {Li , Lj } = Lk , and

i, j, k = x, y, z in cyclic order


2

= L2x

+ L2y

 Li , L2 = 0,

(7.39)

(7.40)

+ L2z .

where L It also follows from the definition of the Poisson bracket that the derivative of a canonical function with respect to generalized coordinates qj is equal to the Poisson bracket of that function with the canonically conjugate momentum pj , that is, ∂F = {F, pj }. ∂qj

(7.41)

In particular, we obtain ∂F = {F, px }, ∂x ∂F = {F, py }, ∂y

(7.42a) (7.42b)

7.3 Postulates of Quantum Mechanics

363

∂F = {F, pz }, ∂z

(7.42c)

F(x + dx, y, z) = F(x, y, z) + {F, px }dx,

(7.43a)

F(x, y + dy, z) = F(x, y, z) + {F, py }dy,

(7.43b)

F(x, y, z + dz) = F(x, y, z) + {F, pz }dz.

(7.43c)

or equivalently,

Thus, the canonical momenta px , py , and pz are called the generators of infinitesimal translations along the x, y, and z directions, respectively. In general, a mechanical description of a physical system requires the concepts of (a) variables or observables, (b) states, (c) equations of motion. Physically, measurable quantities are called observables. In classical mechanics, examples of variables or observables are position, momentum, angular momentum, and energy, which are the characteristics of a physical system. They can be measured experimentally and are represented by dynamical variables, which are well-defined functions of two canonically conjugate variables (generalized coordinates and generalized momenta). So the observables in classical mechanics are completely deterministic. There are states which describe values of the observables at given times. The state of a physical system at a time t = t0 > 0 is uniquely determined by the appropriate physical law and the initial state at t = 0. For example, the state of a system of n interacting particles is determined by assigning 3n position coordinates and 3n velocity coordinates. Finally, there are equations of motion which determine how the values of the observables change in time. As mentioned earlier, Newton’s equations, Lagrange’s equations, or Hamilton’s equations are well-known examples of equations of motion.

7.3 Basic Concepts and Postulates of Quantum Mechanics Classical physics breaks down at the levels of atoms and molecules. Historically, the first indication of a breakdown of classical ideas occurred in the rather complex phenomenon of the so-called black body radiation, which essentially deals with electromagnetic radiation in a container in equilibrium with its surroundings. In other words, the black body radiation is concerned with the thermodynamics of the exchange of energy between radiation and matter. According to principles of classical physics, this exchange of energy is assumed to be continuous in the sense that light of frequency ν can give up any amount of energy on absorption, the exact amount in any particular case depending on the energy intensity of the light beam. Specifically, in 1900 Max

364

Chapter 7

Quantum Mechanics

Planck (Max Karl Ernst Ludwig Planck (1858–1947)) first postulated that the vibrating particles of matter are regarded to act as harmonic oscillators and do not emit or absorb light continuously, but instead only in discrete quantities. Mathematically, the radiation of frequency ν can only exchange energy with matter in units of hν, where h is the Planck constant of numerical value h = 2π
= 6.625 × 10−27 erg sec = 4.14 × 10−21 MeV sec

(7.44)

and
is called the universal constant. Clearly, h has dimension (energy × time) of action, which is a dynamical quantity in classical mechanics. Equivalently, Planck’s quantum postulate can be stated by saying that radiation of frequency ν behaves like a stream of photons of energy E = hν =
ω,

(7.45)

which may be emitted or absorbed by matter where ω = 2π ν is the angular frequency. Clearly, Planck’s constant h measures the degree of discreteness that was required to explain the energy distribution of the black body radiation. Thus, the concept of discreteness is fundamental in quantum mechanics, but it is totally unacceptable in classical physics. Finally, it is important to point out that the Planck equation (7.45) is fairly general so that it can be applied to any quantum system as a fundamental relation between its energy E and the frequency of ν of an oscillation associated with the system. Also, the failure of classical concepts when applied to the motion of electrons appeared most clearly in connection with the hydrogen atom. According to the Rutherford (Ernest Rutherford (1871–1937)) classical model, an atom can be considered as a negatively charged electron orbiting around a relatively massive, positively charged nucleus. With the neglect of radiation, this system is exactly similar to the motion of a planet round the sun, with gravitational attraction between the masses being replaced by the Coulomb attraction between the charges. The potential energy of the Coulomb attraction between the fixed nucleus charge +Ze and the electron of charge −e is V(r) = −Ze2 /r. The hydrogen atom consists of two particles: the nucleus, a proton of mass mp and charge +e (Z = 1), and an electron of mass me and charge −e. The nucleus is small and heavy (mp /me ∼ 2000) and the radius of the proton ∼10−3 times the atomic radius. According to the classical atomic theory of Rutherford, the attractive potential would cause the electron to orbit around the nucleus, and the orbiting electron constitutes a rapidly accelerating charge, which according to Maxwell’s theory acts as a source of radiant energy (James Clerk Maxwell (1831–1879)). Thus, the accelerated charged electron would continuously radiate energy, and in a matter of 10−10 seconds the electron should coalesce with the nucleus, causing the atom to collapse. On the other hand, the frequency of the emitted radiation is related to that of the electron in its orbit. As the electron radiates energy, this frequency, according to classical theory, must change rapidly but continuously, thus giving

7.3 Postulates of Quantum Mechanics

365

rise to radiation with a continuous range of frequencies. Thus, the Rutherford classical atomic model has two important qualitative weaknesses: (a) The atom should be very unstable. (b) It should radiate energy over a continuous range of frequencies. Both of these results are totally contradicted by experiments. The original problem of quantum mechanics was to investigate the stability of atoms and molecules, as well as to explain the discrete frequency spectra of the emitted radiation by excited atoms. The remarkable success in predicting observed atomic and molecular spectra is one of the major triumphs of quantum mechanics. In this chapter, we present the basic principles of quantum mechanics as postulates that will then be used to discuss various consequences. No attempt will be made to derive or justify these postulates. Both the number and content of the basic postulates are to some extent a matter of individual choice. The postulates together with their consequences form a basic but limited theory of quantum mechanics. It has been mentioned in previous sections that classical mechanics identifies the state of a physical system with the values of certain observables (for example, the position x and the momentum p) of the system. On the other hand, quantum mechanics makes a very clear distinction between states and observables. So we begin with the first postulate concerning the state of a quantum system.

Postulate I. (The state vector) Every possible state of a given system in quantum mechanics corresponds to a separable Hilbert space over the complex number field. A state of the system is represented by a nonzero vector in the space, and every nonzero scalar multiple of a state vector represents the same state. Conversely, every nonzero vector in the Hilbert space and its nonzero scalar multiples represent the same physical state of the system. The particular state vector to which the state of the system corresponds at time t is denoted by (x, t) and is called the time-dependent state vector of the system. The state of a physical system is completely described by this state vector (x, t) in the sense that almost all information about the system at time t can be obtained from the vector (x, t). Usually, a state vector is denoted by ψ(x). In the Dirac notation, any general state vector ψ(x) is written as ψ(x) = x|ψ

(7.46)

ψ(x) = ψ|x.

(7.47)

and its complex conjugate as

366

Chapter 7

Quantum Mechanics

This postulate makes several assertions. First, all physical properties of a given system are unchanged if it is multiplied by a nonzero scalar. We can remove this arbitrariness by imposing the normalizing condition    2 ψ(x)ψ(x)dx = ψ(x) dx = 1, (7.48) where the integral is taken over all admissible values of x. Or, equivalently,    2 ψ|xx|ψdx = ψ|x dx = 1, (7.49) or, simply as an abbreviated form, ψ|ψ = 1.

(7.50)

Clearly, the norm of the state vector is unity. In other words, if c is any scalar of modulus unity satisfying |c|2 = 1, then the two state vectors ψ(x) and ϕ(x) = cψ(x) are the same in the sense that they correspond to the same physical state. Clearly, for any two state vectors |ϕ and |ψ, the complex number ϕ|ψ represents the inner product in the space and is given by   ϕ|ψ = ϕ(x)ψ(x)dx = ϕ|xx|ψdx. (7.51) This result combined with (7.46) and (7.47) immediately gives ϕ|ψ = ψ|ϕ.

(7.52)

Two state vectors |ϕ and |ψ are called orthogonal if ϕ|ψ = 0.

(7.53)

A set of state vectors |ψ1 , |ψ2 , . . . is orthonormal if ψi |ψj  = δij . An arbitrary vector |ψ in the state space can be expressed as a sum of a complete set of orthonormal vectors |ψi  in the form  ci |ψi , (7.54) |ψ = i

where the coefficients ci are given by ci = ψi |ψ.

(7.55)

Obviously, this postulate implies the principle of superposition. The number of elements in an orthonormal set of state vectors is called the dimension of the state space. If the dimension is finite, then (7.54) is a finite series. On the

7.3 Postulates of Quantum Mechanics

367

other hand, if the dimension is infinite, (7.54) is an infinite series and hence the question of convergence will arise. Second, Postulate I asserts that all that can possibly be known about the state of the quantum system at time t can be obtained from its state vector (x, t). On the contrary, in classical mechanics, the state of a system at time t is completely described if and only if the state variables q(t) and p(t) are given as definite real numbers. However, the x in (x, t) does not depend in any way on the time variable t. The third assertion of Postulate I is that the functional dependence of  on t is essentially different from its dependence on x. Indeed, as a Hilbert space vector, (x, t) is a function of x alone, and t serves as a parameter. More precisely, (x, t1 ) and (x, t2 ) are to be considered as two different vectors in the space, that is, two different functions of x, which specify the state of the system at two different times t1 and t2 . The state vector (x, t) is often referred to as the state function or the wave function of the quantum system. Naturally, the time evolution of the state function is another important matter, which will be considered later on. In classical mechanics, observables are simply dynamical variables such as position, momentum, and functions of position and momentum. Examples of observables also include angular momentum and energy, which are welldefined functions of two conjugate variables: generalized momentum and generalized coordinates. The measurement (or observation) of a classical observable is a well-defined physical operation that, when performed on the system, yields a single real number, the so-called value of the observable. In the study of classical systems, it is tacitly assumed that the operations of measurement do not significantly disturb their motion. More precisely, any disturbances caused by measurement are negligible. So, in classical mechanics, there is no real difference between the mathematical representation of an observable and the values of the observables. These lead us to formulate the second postulate of quantum mechanics.

Postulate II. (Observable operators and their values) (a) To every physical observable in quantum mechanics, there corresponds in ˆ which has a complete set the Hilbert space a linear Hermitian operator A, of orthonormal eigenvectors {ψn } with the corresponding eigenvalues {λn } such that ˆ n = λn ψn , Aψ

n = 1, 2, 3, . . . .

(7.56)

Conversely, to each such operator in the Hilbert space there corresponds some physical observable. (b) The only possible values of a physical observable are the various eigenvalues. According to this postulate, a quantum observable is mathematically represented by a linear Hermitian operator on an infinite dimensional separable

368

Chapter 7

Quantum Mechanics

Hilbert space and there is one such operator for each quantum observable such as the position, the momentum, the energy, and so on. Such operators in quantum mechanics are called observable operators, which usually satisfy the rules of the theory of operators in the Hilbert space. So the quantum observables are operators in a Hilbert space. However, an observable in classical mechanics is simply a real function of conjugate variables in Euclidean space. This postulate has several remarkable mathematical consequences. The first consequence is that the Hermitian property of an observable operator Aˆ ensures that its eigenvalues {λn } are real. The real eigenvalues correspond to posˆ The secsible measurable random quantities of the observable operator A. ond consequence of Postulate II is that the nature of the representation of the eigenvalues {λn } seems to suggest that they form a discrete set, which leads to the possibility for allowing certain observable operators to be “quantized.” The discreteness is one of the inherent features of quantum phenomena. However, the eigenvalues of an observable operator can be discrete or continuous or a combination of the two depending entirely on the particular observable operator. The most profound difference between these rules of two operators and the rules for manipulating classical observables (or dynamical variables) is that two classical observables A, B always commute (AB = BA), but two observable operators Aˆ and Bˆ do not commute in general. The commutator of these operators is defined by   ˆ Bˆ = Aˆ Bˆ − Bˆ A. ˆ A, (7.57) ˆ This is essentially the difference between operating first with Bˆ and then A, ˆ So, in general, the commutator of two observable and first with Aˆ and then B. operators is nonzero. Physically, the nonzero value must be associated with the magnitude of the unavoidable disturbances between the two measurements of Aˆ and Bˆ in two different orders. It is appropriate to mention here what we mean by a measurement from a simple physical point of view. A measurement is simply a well-defined physical operation which, when performed on a physical system, yields a single real number without any error or ambiguity in the sense that there is no experimental uncertainty associated with the number obtained. If Aˆ and Bˆ represent the measurements of two particular quantum observables (for example, Aˆ might mean measurement of position and Bˆ that of momentum), then Aˆ Bˆ − Bˆ Aˆ = 0. This means that measurement in quantum theory is fundamentally different from that in classical mechanics. A measurement of quantum observables does affect the system, and so it may be expected to appear explicitly in the theory. In other words, it may not be possible to measure two different observable operators of a system simultaneously, and measurement of one may change the value of the other. Physically, observation (or measurement) of something always involves shining light (photons)

369

7.3 Postulates of Quantum Mechanics

on it, which means striking it with photons. If the position is to be determined accurately, the wavelengths of the corresponding waves must be very small, their frequency correspondingly very high, and the momentum of the photons consequently above a certain limit. A blow with such a stream of photons may significantly disturb the observed system if it is small enough. In particular, the position observable, xˆ and the corresponding momentum observable, pˆ = −i
∂/∂x satisfy the celebrated Heisenberg commutation relation [ˆx, pˆ ] = i
.

(7.58)

This follows immediately from the fact that [ˆx, pˆ ]ψ = (ˆxpˆ − pˆ xˆ )ψ = −i
ˆx

∂ψ ∂ψ + i
ˆx + i
ψ = i
ψ. ∂x ∂x

Clearly, the concept of the commutator of two observable operators is of fundamental physical importance and plays a crucial role in the mathematical development of quantum mechanics. Two observables are called complementary if the corresponding operators associated with them do not commute. In one-dimensional problems, the position and momentum are examples of complementary observables. So the simultaneous eigenstates of xˆ and pˆ do not exist. On the other hand, two observables are said to be compatible if measurements of one do not affect the value of the other. So the corresponding operators associated with them must commute. Any two classical observables always commute. In quantum mechanics, energy and momentum are examples of compatible observables. More precisely, if ∂ Aˆ = pˆ = −i
∂x

and

Bˆ = Tˆ =

pˆ 2
2 ∂ 2 =− , 2m 2m ∂x2

then 

  , 
2 ∂ 2 ∂2 ∂ i
3 ∂ ∂ 2 ∂ ˆ ψ =0 − ψ= pˆ , T ψ = −i
, − ∂x 2m ∂x2 2m ∂x ∂x2 ∂x2 ∂x

for any state vector ψ(x). This means that energy and momentum of a free particle can be determined precisely and simultaneously. However, this is not true if the particle moves in a potential field V(x). In this case, pˆ = −i


∂ ∂x

and

ˆ = H

pˆ 2 ˆ x) = Tˆ + V(ˆ ˆ x) + V(ˆ 2m

so that 

       ∂ Vˆ ˆ ψ = pˆ , Tˆ + Vˆ ψ = pˆ , Tˆ ψ + pˆ , Vˆ ψ = −i
ψ = 0. pˆ , H ∂x

370

Chapter 7

Quantum Mechanics

This implies that the momentum and energy of a particle moving in a potential field cannot be determined precisely and simultaneously. The third consequence of Postulate II is that the eigenvectors belonging to different eigenvalues of an observable operator form a complete orthonormal set, that is, for all n and m,  ∞ ψn |ψm  = ψn (x)ψm (x)dx = δmm (7.59) −∞

and any state vector ψ(x) can be expanded in terms of the eigenvectors {ψn } as ∞  ψ(x) = ψn |ψψn (x). (7.60) n=1

In other words, the set of such eigenvectors {ψn (x)} forms an orthonormal basis in the associated Hilbert space. This set {ψn } is often called the eigenbasis ˆ of the observable operator A. Finally, in classical mechanics, both the state and the observable depend on time t. However, no such characterization holds true for the state and the observables in quantum theory. It follows from Postulates I and II that the state (or wave) function (x, t) can change with time t, but the observable operator Aˆ along with its eigenbasis {ψn (x)} and its real eigenvalues {λn } are all independent of time t. If we express the wave function (x, t) in terms of the eigenvectors {ψn (x)} as in (7.60), that is, (x, t) =

∞ 

ψn |ψn (x),

(7.61)

n=1

then the coefficients of this expansion of (x, t) relative to the eigenbasis ψn (x) of Aˆ are time-dependent scalars of the form  ∞ ψn | = ψn (x)(x, t)dx, n = 1, 2, 3, . . . . (7.62) −∞

Such an expansion of the wave function (x, t) in terms of the eigenbasis of the observable operator Aˆ plays a fundamental role in the development ˆ ψn (x), and λn are all indeof quantum mechanics. However, the fact that A, pendent of time does not necessarily mean that the measured values of the observable operator Aˆ will be constant in time. At any rate, it remains to be specified how the result of a particular measurement depends on the state of the system at the time of the measurement. It is then necessary to establish some sort of logical connection between physical observables in classical mechanics and in quantum mechanics. This

371

7.3 Postulates of Quantum Mechanics Observables

Classical quantities

Quantum operators

Position

x or r

xˆ or rˆ

Linear momentum

px = m˙x

∂ pˆ x = −i
∂x

Linear momentum

p = (px , py , pz )

Potential energy

V(r, t)

Kinetic energy

T=

ˆ r, t) V(ˆ
2 Tˆ = − 2m ∇2 
2 ∂ 2 = − 2m ∂x2 +

p2 2m p2

+ V(x)

Hamiltonian

H(x, p) =

Angular momentum

Lx = ypz − xpy

L=r×p

∂ ∂ ∂ pˆ = −i
∇ = −i
∂x , ∂y , ∂z

2m

Ly = zpx − xpz Lz = xpy − ypx

∂2 ∂y2 2

+

∂2 ∂z2



ˆ x, pˆ ) = −
2 ∇ + V(ˆ ˆ x) H(ˆ 2m   ˆLx =
y ∂ − z ∂ i ∂z ∂y  ∂  Lˆ y =
i z ∂x − x ∂z∂   Lˆ z =
x ∂ − y ∂ i

∂y

∂x

Table 7.1 Physical quantities in classical mechanics and the corresponding quantum mechanical operators connection can be made through the celebrated Correspondence Principle, which is the topic of the next postulate.

Postulate III. (Correspondence Principle) A quantum observable operator corresponding to a dynamical variable is obtained by replacing the canonical variable in classical mechanics by the corresponding quantum mechanical operator. This postulate gives a method to formulate the quantum mechanical operators from the classical dynamical variables which are made of the canonical conjugate variables p and q. Like so many other things in quantum mechanics, the Correspondence Principle, first proposed by Niels Bohr (Niels Henrik David Bohr (1885–1962)), is a very deep subject and has many remarkable ramifications. Thus, we have a list of correspondences between physical quantities in classical mechanics and the quantum mechanical operators, as shown in Table 7.1. In general, any observable in classical mechanics is some well-behaved function of position and momentum, f (x, p) and is represented in quantum mechanics by the operator f (ˆx, pˆ ). In one dimension, xˆ = x and pˆ = −i
(d/dx) so that   d f (ˆx, pˆ ) = f xˆ , −i
. dx The position, momentum, angular momentum, and energy are not the only observables in quantum mechanics; there may be several other observables, some of them having no analogues in classical mechanics. For example, quantum observables such as spin and isospin, which characterize some other attributes of the quantum system, have no classical analogue.

372

Chapter 7

Quantum Mechanics

However, this postulate does not say anything about how to specify the quantum observables corresponding to the basic canonical variables. An answer to this point is provided by the following postulate.

Postulate IV. (Quantization) Every pair of canonically conjugate observable operators satisfies the following Heisenberg (Werner Karl Heisenberg (1901–1976)) commutation relations: [ˆqm , qˆ n ] = 0 = [pˆ m , pˆ n ],

(7.63)

[ˆqm , pˆ n ] = i
δˆmn ,

(7.64)

where qˆ m is the observable operator corresponding to the generalized coordinates, and pˆ m is the momentum operator corresponding to the generalized momentum. This postulate is often called the principle of quantization. Clearly, i[pˆ n , qˆ n ] = −
. In Cartesian coordinates, canonical operators xˆ m and pˆ m satisfy [ˆxm , xˆ n ] = 0 = [pˆ m , pˆ n ],

[ˆxm , pˆ n ] = i
δmn ,

where the subscripts m, n take the values 1, 2, 3 for the x, y, z components of xˆ m and pˆ m . Obviously, in one dimension, [pˆ , xˆ ]ψ = (pˆ xˆ − xˆ pˆ )ψ = −i


∂ψ ∂ (xψ) + i
x = −i
ψ ∂x ∂x

so that [ˆx, pˆ ] = xˆ pˆ − pˆ xˆ = i
.

(7.65)

Any observable operator that has a classical analogue can always be expressed as a function of qm and pm , and the quantization rules of such an observable will follow from (7.63) and (7.64). For example, the angular momentum operator Lˆ = (Lˆ x , Lˆ y , Lˆ z ) satisfies the following properties: 

 Lˆ x , Lˆ y = i
Lˆ z ,



 Lˆ y , Lˆ z = i
Lˆ x ,

and



 Lˆ z , Lˆ x = i
Lˆ y .

(7.66)

It is now necessary to establish some sort of mathematical relationship between the state vector and the observable operator of a quantum system. This mathematical relationship is formulated through the concept of measurement. The quantum theory of measurement thus forms the keystone of the mathematical structure of quantum mechanics. According to Postulate II, the only real values that can ever be predicted for an observable are the eigenvalues of the corresponding operator of a given physical system. As to which one of these eigenvalues will be obtained in any given situation, the answer to this

7.3 Postulates of Quantum Mechanics

373

question is expected to be determined by the particular form of a state vector of the physical system at the time of measurement. But the operations of measurement generally affect the system. Therefore, it is almost impossible to predict a single measurement with absolute certainty. Naturally, this difficulty can somehow be resolved by dealing with a large number of measurements under identical conditions and then introducing the average of all results so obtained. This average, called the expectation value of an observable operator, is defined as follows:

Definition 7.3.1.

(Expectation value) ˆ of an observable operator Aˆ in the state ψ(x) of a physThe expectation value A ical system is defined by ˆ $ % ψ|Aψ Aˆ = . ψ|ψ

(7.67)

If the state ψ is normalized, then the expectation value is $ % $  % ˆ . Aˆ = ψ Aψ

(7.68)

ˆ we define the root-mean-square deIn terms of the expectation value of A, ˆ ˆ viation, (∇ A), which measures the dispersion around the mean value A.

Definition 7.3.2. (Root-mean-square deviation) ˆ is defined by the square root of the expectaThe root-mean-square deviation (∇ A) ˆ 2 in the state ψ in which A ˆ is computed. tion value of (Aˆ − A) It follows from the definition that $ %2 % $  $ %2 %  2 $ = ψ  Aˆ − Aˆ ψ ∇ Aˆ = Aˆ − Aˆ $ $ %  $ % %  $ % 2 = Aˆ − Aˆ ψ  Aˆ − Aˆ ψ =  Aˆ − Aˆ ψ 

(7.69)

ˆ is real. because Aˆ is Hermitian and A

Theorem 7.3.3.  2 $ 2 % $ %2 ∇ Aˆ = Aˆ − Aˆ , $ %  2 ˆ . (b) Aˆ 2 = Aψ (a)

(7.70) (7.71)

374

Chapter 7

Quantum Mechanics

Proof: (a) $ %2 % $  $ %2 %  2 $ = ψ  Aˆ − Aˆ ψ ∇ Aˆ = Aˆ − Aˆ $  $ % $ %2  % = ψ  Aˆ 2 − 2Aˆ Aˆ + Aˆ ψ $  % $  %$ % $ %2 ˆ Aˆ + Aˆ ψ|ψ = ψ Aˆ 2 ψ − 2 ψ Aψ $ % $ %2 $ %2 = Aˆ 2 − 2 Aˆ + Aˆ (since ψ|ψ = 1) $ 2 % $ %2 = Aˆ − Aˆ . (b) $

% $  %  2 % $  ˆ Aψ ˆ = Aψ ˆ . Aˆ 2 = ψ Aˆ 2 ψ = Aψ




Theorem 7.3.4. A necessary and sufficient condition for a physical system to be in an eigenstate of an observable Aˆ is ∇ Aˆ = 0.

Proof: The condition is necessary. Suppose the system is an eigenstate ψ of ˆ = λψ. Clearly, Aˆ with a real eigenvalue λ so that Aψ $ % $  % ˆ = λ, Aˆ = ψ Aψ $ 2 %  2 ˆ  = λ2 . Aˆ = Aψ ˆ = 0. Hence, by Theorem 7.3.3(a), (∇ A) The condition ∇ Aˆ = 0 is sufficient. Obviously,   2  $ % 2 0 = ∇ Aˆ =  Aˆ − Aˆ ψ  implies that

or



$ % Aˆ − Aˆ ψ = 0 $ % ˆ = Aˆ ψ. Aψ

ˆ Thus, ψ is an eigenvector of Aˆ with an eigenvalue A.




If ψn (x) is an eigenvector of Aˆ with the corresponding eigenvalue λn so that ˆAψn = λn ψn is satisfied for all n, then the expectation value in the orthonormal eigenstate ψn (x) is given by ˆ n  ψn |λn ψn  $ % ψn |Aψ Aˆ = = = λn . ψn |ψn  ψn |ψn 

(7.72)

7.3 Postulates of Quantum Mechanics

375

ˆ then it can be exOn the other hand, if ψ(x) is not an eigenvector of A, panded in terms of a complete orthonormal set of eigenvectors {ψn (x)}. Consequently, ψ(x) =

∞ 

cn ψn (x) =

n=1

∞ 

ψn |ψψn (x)

(7.73)

n=1

so that ψ 2 =

∞ 

|cn |2 =

n=1

∞    ψn |ψ2 .

(7.74)

n=1

Then the expectation value given by (7.67) assumes the form  ∞ (∞ )    $ % 1  Aˆ = cm ψm A cn ψn  ψ 2 m=1 n=1  ∞ (∞ )   1  cm ψm  cn λn ψn =  ψ 2 m=1

n=1

1  cm cn λn ψm |ψn  = ψ 2 m,n 1  cm cn λn δmn ψ 2 m,n  ∞ 2 2 n |cn | λn n=1 |cn | λn  = = ∞ 2 ψ 2 n=1 |cn |  ∞  ∞   |cn |2 ∞ = = an λn , λ n 2 n=1 |cn | n=1 n=1 =

(7.75) (7.76)

where |cn |2 |ψn |ψ|2  an = ∞ = . ∞ 2 2 n=1 |cn | n=1 |ψn |ψ|

(7.77)

The result (7.76) can be interpreted as the outcome of a large number of ˆ is the measurements under the identical conditions. In other words, A weighted average of a large number of measurements of the system. According to Postulate II, each measurement yields one of the eigenvalues of aˆ belonging to the set {λn }. The numerator of (7.77) represents the frequency with which the eigenvalues {λn } occur in the measurement, while the denominator of (7.77) is the total number of measurements. Hence, the  ratio of the fre2 quency, |ψn |ψ|2 , and the total number of measurements, ∞ n=1 |ψn |ψ| , represents the weight an of the eigenvalues in the measurement. This clearly

376

Chapter 7

Quantum Mechanics

leads to a statistical interpretation of an as the probability that a measurement will yield the eigenvalue λn for the operator Aˆ in the state ψ(x) of the physical system. Furthermore, it follows from (7.77) that the total probability ∞ a n=1 n = 1 reconfirms the statistical interpretation, since an is the probabilˆ ity that a single measurement yields an eigenvalue of A. For a normalized state vector ψ, ψ 2 =

∞ 

 2 |cn |2 = ψn |ψ = 1,

(7.78)

n=1

 2 an = |cn |2 = ψn |ψ .

(7.79)

Consequently, (7.76) reduces to ∞   $ %  ψn |ψ2 λn . ˆA =

(7.80)

n=1

Thus, the quantity |cn |2 = |ψn |ψ|2 can be recognized as the probability that the measurement will yield the eigenvalue λn of Aˆ in the normalized state ψ(x). Clearly, 0 ≤ |ψn |ψ|2 ≤ 1. All these results lead us to Postulate V, which is essentially concerned with the outcome of a measurement of a quantum observable of a physical system in a given state.

Postulate V. (Outcome of quantum measurement) If an observable operator

Aˆ has eigenbasis {ψn } with the corresponding eigenvalues {λn }, then the probability that the measurement will yield the eigenvalue λn of Aˆ of the system in the normalized state ψ(x) is  2 P(λn ) = ψn |ψ . (7.81)

This postulate makes several radical assertions. First, it is impossible to predict with absolute certainty the outcome of a measurement that is made on a quantum system in a completely defined state. This is because of the fact that any measurement entails an interaction between the measuring equipment and the measured system. Thus, the prediction of a quantum measurement is of statistical nature. This is perhaps the most radical feature of the quantum theory of measurement and completely contrary to the spirit of classical mechanics, which is totally based on the deterministic principle. Second, when a quantum system is in a state ψ(x), the value obtained in measurement of an observable operator Aˆ is a random variable with a probability distribution whose mean value is associated with a large number of measurements in this particular state. This leads to a statistical interpretation of quantum measurements. Third, if two repeated measurements of some observable are performed

7.4 The Heisenberg Uncertainty Principle

377

on a system in a given state, the results of these two measurements will not necessarily be the same, since the operations of measurements affect the system in general. In classical mechanics, two or more repeated measurements of a classical observable will always produce identical results. So this postulate makes the most remarkable difference between quantum and classical mechanics because of the special role played by the process of measurement in quantum mechanics. In the quantum theory of measurement, it is important to determine the degree to which a random measurement deviates from the expectation value. The standard deviation, which is a measure of the dispersion around the expectation value, is usually called the uncertainty Aˆ (or root-mean-square deviation) introduced in Definition 7.3.2.

7.4 The Heisenberg Uncertainty Principle If Aˆ and Bˆ are Hermitian observables, the uncertainties Aˆ and Bˆ in the measurement of Aˆ and Bˆ in the given state ψ(x) are given by $ %2 %1/2 $ Aˆ − Aˆ $  $ %2 %1/2 = ψ  Aˆ − Aˆ ψ $ % %1/2 $ %  $ = Aˆ − Aˆ ψ  Aˆ − Aˆ ψ  $ %  =  Aˆ − Aˆ ψ  = ψ1 ,

Aˆ =

where

(7.82)

 $ % ψ1 = Aˆ − Aˆ ψ.

(7.83)

Bˆ = ψ2 ,

(7.84a)

 $ % ψ2 = Bˆ − Bˆ ψ.

(7.84b)

Similarly, we write where

Since the state vector ψ is an element of a Hilbert space, there exists a correlation between the uncertainties. The correlation is well known as the Generalized Uncertainty Principle.

Theorem 7.4.1. (The Uncertainty Principle) If Aˆ and Bˆ are Hermitian operators, then for any state vector   1  1 $ ˆ % ˆ ˆ ˆ AB >  A, B . 2 i

(7.85)

378

Chapter 7

Quantum Mechanics

Proof: We have Aˆ = ψ1 , Bˆ = ψ2 ,

(7.86)

$ %  ψ1 = Aˆ − Aˆ ψ,  $ % ψ2 = Bˆ − Bˆ ψ.

(7.87)

According to Schwarz’s inequality, we have, for any two vectors ψ1 and ψ2 ,   ψ1 ψ2 ≥ ψ1 |ψ2    ≥ Imψ1 |ψ2    1
  ≥  ψ1 |ψ2  − ψ1 |ψ2   2i   1
  ≥  ψ1 |ψ2  − ψ2 |ψ1  . 2i Evidently,  *   1 $ $ %  $ % % $ $ %  $ % %=    ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ AB ≥  A − A ψ B − B ψ − B − B ψ A − A ψ  2i    1 $$   %%   ˆ ˆ ˆ ˆ ≥ ψ AB − BA ψ  2i   1  1 $ ˆ % ˆ ≥  A, B . 2 i




Corollary 7.4.2. For conjugate operators qj and pk , the following holds
qj pk ≥ δjk , 2
xj pj ≥ , 2
, xj vj ≥ 2m

(7.88a) (7.88b) (7.88c)

where vj is the velocity vector and m is the mass.

Proof: Substituting Aˆ = qj and Bˆ = pk in (7.85), we obtain (7.88a). In Cartesian coordinates, (7.88a) reduces to (7.88b). Finally, with the velocity vector vj and mass m, (7.88b) becomes (7.88c).


7.5 The Schrödinger Equation of Motion

379

It is important to point out that (7.88b) is the famous uncertainty relation between the position and momentum first discovered by Heisenberg. According to the Heisenberg uncertainty relation, the position and momentum of a particle cannot be determined exactly and simultaneously. In other words, it is impossible to design an experiment that can measure the precise values of both the position and momentum of a particle. Quantum mechanics is essentially built upon the Heisenberg Uncertainty Principle, and hence this principle is perhaps the most fundamental result in quantum theory. Finally, since
is very small,
/2m is extremely small for any macroscopic system, and hence (7.88c) reveals why the Uncertainty Principle is of no importance in classical mechanics. So when the system is large,
is negligible and all observable operators commute with each other. Consequently, all observables can then be represented by ordinary algebraic variables and can be measured without any mutual disturbance. According to the Corresponding Principle, the definitive relations between these variables, such as position, energy and momentum, are the same as in classical mechanics in the limit
→ 0. So the distinction between quantum and classical systems is not made on the sole basis of spatial extension, but in units of Planck’s constant h. Evidently, the size of a physical system becomes important and must be considered. Thus, for example, for an electron in an atom the typical “action” (length × momentum or time × energy) is the product of the Bohr radius 10−8 cm, and the momentum in the Bohr orbit 10−19 so that their product is exactly of the order of h, so quantum effects are absolutely essential. Thus, classical mechanics can be considered as the limit
→ 0 of quantum mechanics.

7.5 The Schrödinger Equation of Motion In our review of classical mechanics in Section 7.2, the problem of the time evolution of the classical state has been discussed in some detail. It follows from the review that there is a very close connection between energy and time as established in the Hamiltonian formulation of classical mechanics. In fact, it is the energy written as the Hamiltonian function H(x, p) that really governs the time evolution of the classical systems. On the other hand, our previous discussion reveals that there is a remarkable distinction between the state of a system and the physical observables in quantum mechanics. With a reasonably good understanding of these concepts, it is not too difficult to comprehend the time-evolution problem in quantum mechanics. Before we actually embark on the problem, it is necessary to preview some other related ideas of mechanics. In general, any dynamical quantity that varies in space and time constitutes a wave that can typically be represented by the wave function    (r, t) = A(r) exp −i 2π νt + ϕ(r) ,

(7.89)

380

Chapter 7

Quantum Mechanics

where A is the amplitude, ν is the frequency, and ϕ is the phase of the wave. This can also be expressed in the form

where

(r, t) = ψ(r)e−2π iνt ,

(7.90)

  ψ(r) = A(r) exp −iϕ(r) .

(7.91)

According to the Planck equation (E = hν), if a quantum system has energy E, its state vector |(r, t) at time t should contain a factor exp(−2π iνt) = exp[−(iE/
)t] so that   % % (r, t) = e−iEt/
(r, 0) . (7.92) Clearly, energy is an observable, and hence, for a system to have a definite energy E, it must be in an eigenstate of this observable. If this is the case, Equation (7.92) expresses the fact that the state vector at time t is different from that at t = 0 only by a scalar factor, and so it describes the same physical state. For this reason an eigenstate of energy is called a stationary state of the system. The following postulate of quantum mechanics is concerned with the existence of the energy operator and the time development of a quantum system.

Postulate VI. (a) (Hamiltonian operator) For every physical system there exists a linear Herˆ the so-called Hamiltonian operator which represents the mitian operator H, observable operator corresponding to the total energy of the state vector of the system. (b) (Schrödinger’s equation) If a physical system is not disturbed by any experˆ determines the time development of the iment, the Hamiltonian operator H state vector of the system (r, t) through the partial differential equation i


∂ ˆ = H(r, t). ∂t

(7.93)

This is called the time-dependent Schrödinger equation, and represents the fundamental equation of motion in quantum mechanics first discovered by Erwin Schrödinger (1887–1961). Its ultimate justification is that it leads to predictions which are in remarkable agreement with experimental findings. As in equations of motion in classical mechanics, Equation (7.93) is completely deterministic in the sense that, given the state (r, t) at some time t = t0 , Equation (7.93) will uniquely determine the state (r, t) at some other time t. ˆ corresponds to the force in classical mechanics The Hamiltonian operator H ˆ includes the potential energy which gives the force because the total energy H ˆ is equivalent to the force field. in classical mechanics. So H

7.5 The Schrödinger Equation of Motion

381

For a single particle of mass m moving in space, the classical state is described by the position and momentum vectors (r, p). If the particle is under the action of a force F(r) which is derived from a potential V(r) so that F = −∇V , the Hamiltonian function H is H(r, p) = T + V =

p2 + V(r). 2m

(7.94)

ˆ corresponding to this classical function H The Hamiltonian operator H in quantum mechanics is derived by replacing r and p with the operators rˆ and pˆ = −i
∇, respectively. Consequently, the Schrödinger equation (7.93) assumes the form , ∂
2 ˆ i
= − ∇ 2 + V(ˆr)  = H, (7.95) ∂t 2m where ˆ =− H


2 2 ∇ + V(ˆr) 2m

(7.96)

and (r, t) belongs to the Hilbert space L2 (R3 ). This postulate has three main consequences. First, it asserts that (7.92) can ˆ be derived as a solution of the Schrödinger equation (7.93). We assume that H has a purely discrete spectrum of eigenvalues so that it has a complete set of eigenstates |ψn (r) with the corresponding eigenvalues En . Then |(r, t) can be expanded in terms of the complete eigenstates as   %  % (r, t) = an (t)ψn (r) .

(7.97)

n

Substituting this into (7.93) gives i


 dan dt

|ψn  =



En an (t)|ψn ,

n

which, equating the coefficients of |ψn , yields i


dan = En an (t). dt

(7.98)

Hence, the solution of this equation is   iEn t , an (t) = an (0) exp −


(7.99)

382

Chapter 7

Quantum Mechanics

where an (0) is the initial value of an (t) at t = 0. Thus, the time-dependent solution is    %  % iEn t  (r, t) = an (0) exp − (7.100) ψn (r) .
n The second consequence of (7.93) is the de Broglie wave (Louis Victor Pierre Raymond duc de Broglie (1892–1987)) for free particles of definite momentum p =
k, which is required to explain electron diffraction. The onedimensional de Broglie wave is x|pa  = aeikx = aeipx/
,

(7.101)

where a is a constant. If it is normalized by using the orthonormality condition for continuous eigenvalues α|α  = δ(α − α ), where the simplest representation of the Dirac delta function is  ∞ 1 δ(α) = eiαx dx, 2π −∞ then



p|p  =

p|xx|p dx 

= |a|2

∞ −∞

  exp −i(p − p )x/
dx

= δ(p − p )

(7.102)

so that |a|2 = (2π
)−1 . In the case of a free particle of definite momentum, 1 eipx/
. x|ψ = x|p = √ 2π
According to (7.101), the time-dependent solution of (7.93) is    % (x, t) = √ 1 exp i(px − Et)/
, 2π


(7.103)

where E = (p2 /2m). This is the complete de Broglie wave. It appears now for the first time as a consequence of the general quantum mechanical formalism and is direct evidence for the correctness of the postulated Schrödinger equation of motion. The third consequence of the Schrödinger equation is the conservation law of total probability in space and time. The Schrödinger equation (7.95) is satisfied by the complex wave function of real variables so that its real and imag-

7.5 The Schrödinger Equation of Motion

383

inary parts separately satisfy (7.95). The complex conjugate equation is obtained by changing the sign of the imaginary part everywhere so that −i


∂
2 = − ∇ 2  + V(r). ∂t 2m

If the general solution of (7.95) has the form ,  iEn t (r, t) = an ψn (r) exp − ,
n

(7.104)

(7.105)

then its complex conjugate satisfies (7.104) and has the representation ,  iEn t (r, t) = an ψ n (r) exp . (7.106)
n Then the time rate of change of the probability density P(r, t) =  = ||2 at any point is  ∂ ∂ ∂ + ,  =  ∂t ∂t ∂t which is, by (7.95) and (7.104),    ∂ i
  2   = −  ∇  −  ∇ 2 ∂t 2m     i
=− div  ∇ − (∇) 2m = − div J, or, equivalently, ∂P + div J = 0, (7.107) ∂t where the vector J may be called the probability current density or the probability flux defined by   i
  (7.108) J=  ∇ − (∇) . 2m Equation (7.107) represents the conservation of total probability in space and time, that is, it neither created nor destroyed. Equivalently, any net increase (or decrease) of probability within a region in space is associated with a net inflow (or outflow) or probability. If the system is in a stationary state, then ∂P/∂t = 0 and hence div J = 0, that is, the probability flow is solenoidal. It is important to point out that the derivation of the equation of continuity for probability flow (7.107) has been based upon the real potential function. The

384

Chapter 7

Quantum Mechanics

complex potential energy function is physically unrealistic because it leads to complex energy eigenvalues. However, it has been utilized extensively in modeling nuclear scattering interactions. Equation (7.107) is identical with the equation of continuity, which represents the conservation of mass in classical hydrodynamics and the conservation of charge in classical electrodynamics provided there are no sources or sinks in the system. The interpretation of the probability flux J as the probability current density leads to the representation of momentum by the operator −i
∇ and velocity by the operator (
/im)∇. It is evident from the definition (7.108) of J that  
J(r, t) = Re  ∇ , (7.109a) im or

 
J(r, t) = Im  ∇ . m

(7.109b)

From the definition, the probability density P(r, t) = |(r, t)|2 is determined by modulus squared of the wave function (r, t). This means that P(r, t)dr is the probability of finding the particle in its volume element dr = dxdydz about its position r at time t, when a large number of precise measurements are made on independent particles each of which is described by the single particle wave function (r, t). The normalizing condition ensures that the total probability of finding the particle somewhere in the space must be unity, so that for each fixed t,    (r, t)2 dr = 1, | = (7.110) where the integral is taken over the entire region of the space. A general statement of the boundary condition on the wave function is needed so that it can be normalized in principle. This probability interpretation of the wave function was first introduced by Max Born (1882–1970) in 1926 and plays a fundamental role in the development of quantum mechanics. The fourth consequence of the postulate is that the time evolution of the state vector as described by Equation (7.93) is consistent with the requirement that the state vector always has a unit norm. If (x, 0) = 0 and 0 |0  = 1, it can be proved that, for all t, d | = 0. dt We have

 2 1  1 2  ∂ ∂  d  | =  +  dt ∂t  ∂t

(7.111)

7.5 The Schrödinger Equation of Motion

 %1$  % 1$ ˆ  ˆ H  H by (7.93) i
i
%$  % 1 $  ˆ ˆ =  H  H = 0, i


385

=

(7.112)

ˆ is Hermitian. This means that  = constant. But 0 = 1, and since H hence the constant is equal to 1. The fifth consequence of Postulate VI(b) is that (7.93) leads to the analogue of the Hamilton–Jacobi equation in classical mechanics. Equation (7.93) can also be rewritten as   ∂
ˆ = 0, ln  + H (7.113) ∂t i or, equivalently, ∂S ˆ = 0, +H ∂t

(7.114)

where S = (
/i) ln  is called Hamilton’s principal function and may be interpreted as the action so that  = exp((i/
)S). So (7.114) is the analogue of the Hamilton–Jacobi equation. A careful interpretation of this postulate further reveals how the states described by the wave functions and the observables represented by the Hermitian operators are directly related to actual measurements. Conversely, the question is how can information obtained about a physical system be incorporated in the wave functions. If an observable Aˆ of a quantum system is measured in a state |(r, t), then Postulate II determines the value λ as an eigenˆ If the same measurement of Aˆ is repeated in a sufficiently short value of A. time interval, then the same value of λ is expected. However, if the time interval between the two measurements is long enough, then the system is expected to change appreciably due to the time variation according to the Schrödinger equation. Consequently, the eigenvalue λ of Aˆ will be different. However, if the observable Aˆ is a constant of motion, it will retain the same value λ obtained by the first measurement. Thus, this case is very much similar to that in classical mechanics where the time evolution of a classical observable is uniquely determined by the classical equations of motion. This discussion also leads to the question of simultaneous measurement of all observables in a quantum system. In general, quantum observables are mutually incompatible in the sense that measurements of one of them partially or completely destroy any knowledge or information about the others. So it is not possible to determine the simultaneous values of all observables of a quantum system.

Example 7.5.1. We shall discuss the plane wave solution of the time-dependent Schrödinger equation for a particle under the action of a constant poten-

386

Chapter 7

Quantum Mechanics

tial. Obtain the solution of a free particle with the initial condition (r, 0) = 0 (r). The time-dependent Schrödinger equation for a particle under the potential V(r) is ,
2 2 (7.115) ∇ . i
t = V(r) − 2m With a constant potential V(r) = constant = V , a plane wave solution of the form   (r, t) = A exp i(κ · r − ωt) is possible provided the dispersion relation ω
= V +


2 κ 2 2m

(7.116)

is satisfied by frequency ω and wavenumber vector κ. The group velocity of the wave is given by Cg = ∇κ ω =


κ . m

(7.117)

Through the Correspondence Principle,
ω is to be interpreted as the total energy,
2 κ 2 /2m as the kinetic energy,
κ as the particle momentum, and hence the group velocity as the classical velocity. We next solve the Schrödinger equation for a free particle (V ≡ 0) with the initial condition (r, 0) = 0 (r),

(7.118)

where 0 (r) has a compact support in R . ˜ Apply the Fourier transform (κ, t) of (r, t) defined by 3

1 ˜ (κ, t) = (2π )3/2

 ∞ 

e−iκ·r (r, t)dr,

−∞

where κ = (k, l, m). Then the Schrödinger equation and the initial condition become ˜t = i



2 κ 2 ˜ , 2m

˜ ˜ 0 (κ). (κ, 0) = 

Hence, the transformed solution of this system is , i
2 ˜ ˜ (κ, t) = 0 (κ) exp − κ t . 2m

(7.119)

7.5 The Schrödinger Equation of Motion

387

The inverse Fourier transformation gives 1 (r, t) = (2π )3/2 1 = (2π )3/2

 ∞ −∞

 ∞ −∞

,
κ 2 t ˜ dκ 0 (κ) exp i κ · r − 2m ,  κ · r
κt ˜ 0 (κ) exp iκ − dκ. (7.120) κ 2m

Thus, the solution (r, t) is a continuous superposition of functions of the form ,  κ · r
κ − t . (7.121) exp iκ κ 2m Such a function represents a plane wave traveling with velocity
κ/2m in a direction normal to the plane κ · r = 0 and having a wavelength 2π/κ. If 0 (κ) is zero outside a small region about a fixed point κ0 = (k0 , l0 , m0 ), then the x, y, and z components of momentum p0 are certain to have values very near to
k0 ,
l0 , and
m0 , respectively, and (r, t) is a superposition of plane waves with wavenumbers κ0 = [(k02 + l02 + m20 )]1/2 = p0 /
, where p0 = [(
k0 )2 + (
l0 )2 + (
m0 )2 ]1/2 is the total momentum. More precisely, in a state where the momentum is certainly very close to p0 , the state function is a superposition of plane waves with wavenumbers close to the value p0 /
. In this sense, a free particle of momentum p is associated with a wave of wavenumber κ0 = p0 /
(or of wavelength λ0 = 2π
/p0 ). In the one-dimensional case, the solution takes the form    ∞   1
˜ 0 (k) exp i kx − γ tk2 dk (x, t) = √  γ= 2m 2π −∞  ∞  ∞   1 = ψ0 (α)dα exp i kx − kα − γ k2 t dk 2π −∞ −∞ ,    ∞  ∞ x − α 2 i(x − α)2 1 ψ0 (α)dα exp −iγ t k − + dk = 2π −∞ 2γ t 4γ t −∞ ,  ∞ i(x − α)2 1−i dα, (7.122) ψ0 (α) exp = √ 4γ t 2 2π γ t −∞ ∞ √ 2 where the last equality follows from the equality −∞ e−iαx dα = π/2α × (1 − i). The integral in (7.122) would be convergent for a suitably chosen function ψ0 (x).


Example 7.5.2. (The free particle de Broglie wave and the wave–particle duality) The interference phenomena, of which diffraction is typical, require the

388

Chapter 7

Quantum Mechanics

wave picture for a satisfactory explanation. The typical wave is given by   (r, t) = a exp i(κ · r − ωt) , (7.123) where a is the amplitude, κ is the wavenumber vector, and ω is the frequency representing the fundamental physical characteristics of the wave. The wave propagates with a constant velocity of light c = ω/|κ|. According to the famous Planck hypothesis, the radiation of frequency ω behaves like a stream of photons of energy E =
ω which may be emitted or absorbed by matter. Since they travel with the velocity of light, according to the theory of special relativity, their rest mass must be zero. The relativistic equation for energy E and momentum p is  2 2 E /c = p2 + m2 c2 . (7.124) Thus, for photons (m = 0), p = E/c. A simple elimination of c gives E =
ω,

(7.125a)

p =
κ.

(7.125b)

This shows clearly the basic relation between the particle parameters (E, p) of the photons, and the wave parameters (ω, κ) of the corresponding wave. It is well known from diffraction experiments with electrons that a beam of electrons traveling with a definite momentum of magnitude p in the x-direction is directly associated with the wavelength of the corresponding wave so that p =
κ. According to Max Born’s physical interpretation of (r, t), |(r, t)|2 represents the density of electrons, that is, the number of electrons per unit volume at (r, t). De Broglie made a remarkable hypothesis that the result (7.125a) and (7.125b) which relates the wave and the particle aspects of radiation should apply also to electrons. Consequently, an electron of given energy and momentum must be associated with the de Broglie wave , i (r, t) = a exp (p · r − Et) . (7.126)
So we can identify the wave function of a free particle of mass m in the form ψ(r) = aeiκ·r ,

(7.127)

where the wavenumber κ and the momentum p are related by (7.125b). Clearly −i
∇ψ = pψ,  2
2 2 p ψ. − ∇ ψ= 2m 2m

(7.128) (7.129)

7.5 The Schrödinger Equation of Motion

389

These results show that the wave function is a simultaneous eigenfunction of momentum and energy with eigenvalues p and p2 /2m, respectively, for all values of p. This is due to the fact that for a free particle the total energy H is a function of the momentum only (H = p2 /2m) as in classical mechanics. The eigenfunctions are mutually orthogonal which follows from the Dirac formalism for eigenvalue problems with continuous spectra:   iκ ·r iκ·r 

ae , ae = |a|2 ei(κ −κ)·r dr = (2π )3 |a|2 δ(κ − κ ). (7.130)


Example 7.5.3. (Momentum wave function ψ(p) as the three-dimensional Fourier transform of ψ(r)) The wave function ψ(r) = ψ(x, y, z) is the solution of the time-independent Schrödinger equation ∇ 2ψ +

2m (E − V)ψ = 0,
2

(7.131)

where E is the total energy and V(x, y, z) is the potential of the field in which a particle moves. To avoid the frequent repetition of the constant 2m/
2 in the equation, atomic units are used. If we use the electronic charge e, the electronic mass m, and the length a = (
2 /me2 ) as the units of charge, mass, and length, respectively, the Schrödinger equation (7.131) becomes ∇ 2 ψ + 2(E − V)ψ = 0,

(7.132)

and the units of energy and velocity are now e2 /a and e2 /
= c/137, respectively, where
c/e2 = 137 and c is the speed of light. ˜ According to the Dirac transformation theory, ψ(p) is the three-dimensional Fourier transform of the wave function ψ(r) provided atomic units are used: 1 ˜ ψ(p) = (2π )3/2

 ∞ 

e−ip·r ψ(r)dr.

−∞

It follows from the Parseval relation of Fourier transforms that  ∞ 

  ψ(r)2 dr = 1 =

−∞

 ∞ 

  ψ(p) ˜ 2 dp,

−∞

˜ where ψ(r) is normalized to unity and hence ψ(p) is automatically normalized to unity. The momentum density function I(p) is defined by  ∞   ψ(p) ˜ 2 p2 d, (7.133) I(p) = 0

390

Chapter 7

Quantum Mechanics

where d is the element of solid angle for p. Physically, I(p)dp can be interpreted as the probability that the magnitude of the momentum of the electron lies between p and p + dp so that the total probability is unity, that is,  ∞ I(p)dp = 1. (7.134) 0

In the Schrödinger interpretation, the momentum p = −i∇ψ in the sense that the mean value of the x-component of p is given by the equation  ∞  px  = −i

ψ ∗ (r)

−∞

∂ψ(r) dr. ∂x

(7.135)

˜ Since F {−i∂ψ/∂x} = px ψ(p),  ∞  px  =

˜ px ψ˜ ∗ (p)ψ(p)dp.

(7.136)

−∞

Similarly, the mean momentum is given by  ∞ p = I(p)dp.

(7.137)

0

For a single electron in a hydrogen-like atom of nuclear charge Z, we choose the Coulomb potential Z V(r) = − , r 2 2 2 2 where r = x + y + z , and for the hydrogen atom Z = 1, for ionized helium Z = 2, and for doubly ionized lithium Z = 3. It can be readily verified by direct substitution in equation (7.132) that  3 1/2 Z ψ(r) = e−Zr (7.138) π is a solution of (7.132) with E = − 12 Z 2 . The constant involved in the solution for ψ(r) is chosen so that ψ(r) is normalized to unity. The momentum wave function associated with this wave function is ˜ ψ(p) =

=



Z3 π

1/2

1 (2π )3/2

 ∞ 

e−Zr−i(p·r) dr

−∞

5 1/2

(8Z ) , π(p2 + Z 2 )2

(7.139)

391

7.5 The Schrödinger Equation of Motion

so that  2 I(p) = 4π p2 ψ(p) =

32p2 Z 5 . π(p2 + Z 2 )4

(7.140)

Hence, the mean momentum is  ∞ p = pI(p)dp 0

32Z 5 = π

,

∞ 0

pdp − Z2 (p2 + Z 2 )3



∞ 0

pdp 8Z . (7.141) = (p2 + Z 2 )4 3π

In terms of the unit of momentum mc/137, the mean momentum has therefore the value 8mcZ/411π . In the nomenclature of spectroscopy, the wave √ function ψ(r) = (1/ π) exp(−r) is the 1s wave function of a hydrogen atom. So, in a 1s orbit, the mean momentum of the electron is 8mZe2 /3π h because the fine structure constant is 137 =
c/e2 .


Example 7.5.4. (Motion of a wave-packet in a field of force) We consider a three-dimensional wave-packet (r, t) and calculate the x-component of its velocity given by    d d ∂ ∗ ∂ x = xdτ +  ∗ x dτ,  ∗ xdτ = dt dt ∂t ∂t where the integration is taken over the whole space. We next substitute for ∂/∂t and ∂ ∗ /∂t from the Schrödinger equation (7.95) and its complex conjugate to obtain ,   2 ∗  2  d i
∗ x = − ∇  xdτ −  x ∇  dτ . (7.142) dt 2m Then we apply Green’s theorem         2 ∗  ∇ ∗ x −  ∗ (∇x) ds, ∇  x −  ∗ ∇ 2 x dτ = where the right-hand side represents the surface integral which is taken over an infinitely distant bounding surface. Since the wave-packet represents a localized particle, it must vanish all over the above surface. Consequently, the surface integral vanishes and hence    2 ∗   ∇  xdτ =  ∗ ∇ 2 x dτ. We substitute this result into (7.142) to obtain   d i
 ∗  2  x = −  ∇ x −  ∗ x∇ 2  dτ dt 2m

392

Chapter 7

Quantum Mechanics

,  i
∂ ∗ 2 2 − x∇  dτ =−  x∇  + 2 2m ∂x ,  1 1 ∗
∂ =  dτ = px . m i ∂x m

(7.143)

Similarly,  d ∂
d px  = dτ ∗ dt i dt ∂x ,   
∂ ∗ ∂ ∂ ∂ = dτ +  ∗ dτ i ∂t ∂x ∂x ∂t    2   
2 2 ∗
2 ∗ ∂ ∗ ∂ = − ∇  + V dτ +  ∇  − V dτ 2m ∂x ∂x 2m  ,  
2  ∗ 2 2 ∗ ∗ ∗ ∂ (V) dτ. V x −  =  ∇ x − x ∇  dτ + 2m ∂x The first integral can be transformed into a surface integral by Green’s theorem. It is assumed that this integral vanishes when taken over a very large surface. Consequently, ,  d ∂ ∂ ∗ px  =  V − (V) dτ dt ∂x ∂x   1 2  ∂V ∂V ∗ =  − dτ = − . (7.144) ∂x ∂x This result is called Ehrenfest’s theorem (Paul Ehrenfest (1880–1933)).




Example 7.5.5. (The three-dimensional time-independent Schrödinger equation in a spherically symmetric potential) We solve the Schrödinger equation (7.131) for a spherically symmetric potential V(r) = V(r). This equation may be written in spherical polar coordinates (r, θ, ϕ) in the form:       ∂ 2ψ 1 ∂ 2 ∂ψ 1 ∂ ∂ψ 1 + K E − V(r) ψ = 0, r + sin θ + 2 2 2 2 2 r ∂r ∂r r sin θ ∂θ ∂θ r sin θ ∂ϕ (7.145) where ψ = ψ(r, θ, ϕ), K = 2m/
2 , 0 ≤ r < ∞, 0 ≤ θ ≤ π , and 0 ≤ ϕ ≤ 2π . We seek a separable solution of the form ψ = R(r)Y(θ, ϕ) and then substitute into (7.145) to obtain two equations ,   d 2 dR r + K(E − V)r2 − λ R = 0, dr dr

(7.146)

,

7.5 The Schrödinger Equation of Motion   ∂ 1 ∂2 1 ∂ sin θ + 2 Y + λY = 0, sin θ ∂θ ∂θ sin θ ∂ϕ 2

393 (7.147)

where λ is a separation constant. We first solve (7.147) by separation of variables through Y(θ, ϕ) = (θ ) × (ϕ) so that this equation becomes     d d sin θ + λ sin2 θ − m2  = 0 sin θ dθ dθ

(7.148)

d2  + m2  = 0, dϕ 2

(7.149)

and

where m2 is the second separation constant. The general solution of (7.149) is  = Aeimϕ + Be−imϕ , where the arbitrary constants A and B are to be determined from the boundary conditions on ψ(r, θ, ϕ) = R(r)(θ )(ϕ). According to the basic postulate of quantum mechanics, the wave function for a particle without spin must have a definite value at every point in space. In particular, ψ must have the same value whether the azimuthal coordinate ϕ is ϕ or ϕ + 2π , that is, (ϕ) = (ϕ + 2π ). Consequently, the solution for  has the form  = Ceimϕ ,

m = 0, ±1, ±2, . . . ,

(7.150)

where C is an arbitrary constant. To solve equation (7.148), it is convenient to change the variable x = cos θ, (θ ) = u(x), −1 ≤ x ≤ 1, so that it becomes  , -   du m2 d  u = 0. 1 − x2 + λ− dx dx 1 − x2

(7.151)

When m = 0, this is the Legendre equation, which gives the Legendre polynomials Pl (x) of degree l as solutions provided λ = l(l + 1), where l is a positive integer or zero. When m = 0, (7.151) with λ = l(l + 1) admits solutions which are well known as the associated Legendre functions Plm (x) of degree l and order m defined by  m/2 dm Plm (x) = 1 − x2 Pl (x), dxm

x = cos θ.

(7.152)

394

Chapter 7

Quantum Mechanics

Clearly, Plm (x) vanishes when m > l. As for negative integral values of m, it can be readily shown that Pl−m (x) = (−1)m

(l − m)! m P (x). (l + m)! l

(7.153)

Hence, the functions Pl−m (x) differ from Plm (x) by a constant factor, and as a consequence, m can take only non-negative integral values. Thus, the associated Legendre functions Plm (x) with |m| ≤ l are the only nonsingular and physically acceptable solutions of (7.151). Since |m| ≤ l, when l = 0, we have m = 0; when l = 1, we have m = −1, 0, +1; when l = 2, we have m = −2, −1, 0, 1, 2; and so on. This means that, given l, there are exactly (2l + 1) different values of m = −l, . . . , −1, 0, 1, . . . , l. The numbers l and m are called the orbital quantum number and the magnetic quantum number, respectively. It is convenient to write down the solutions of (7.147) as functions which are normalized with respect to an integration over the whole solid angle. They are called spherical harmonics and are given by , Ylm (θ, ϕ) =

(2l + 1) (l − m)! 4π (l + m)!

-1/2 (−1)m eimϕ Plm (cos θ ),

m ≥ 0, (7.154)

where Plm (cos θ ) = sin|m| (θ )

d|m| Pl (x), dx|m|

x = cos θ,

(7.155)

and  −1 dl  2 l Pl (x) = 2l l! − 1 . x dxl

(7.156)

The identity (7.156) is known as the Rodrigues formula for the Legendre polynomial Pl (x) of degree l. To obtain the complete solution of (7.145), we combine the spherical harmonic solutions (7.154) with the solutions of the radial equation (7.146). For a given λ = l(l + 1), (7.146) possesses radial eigenfunctions Rn,l (r) in the normalized form corresponding to energy eigenvalues En,l of the system. These energy levels may form a discrete spectrum or a continuous spectrum depending on the nature of the given potential V(r). Finally, the solutions of the eigenvalue problem (7.145) are of the form ψn,l,m (r, θ, ϕ) = Rn,l (r)Ylm (θ, ϕ),

(7.157)

where n is usually called the principal quantum number, which can have the values 1, 2, 3, . . . .


7.6 The Schrödinger Picture

395

7.6 The Schrödinger Picture State vectors in quantum mechanics describe values or probabilities for physical observables at a given time. As in classical mechanics, equations of motion in quantum mechanics determine how these values or probabilities change in time. It is sufficient for a quantum state to specify expectation values of observable operators, and hence the equations of motion are required to describe the time evolution of the expectation values. The expectation value of an observable operator in the state represented by the normalized state vector |(r, t) is given by $ % $ % $  % ˆ . Aˆ = Aˆ  =  A (7.158) The time development of the expectation value can therefore be described in at least three different ways: (a) The state vector (r, t), which will be simply written as (t), changes with time t, and each observable quantity is represented by an operaˆ which is independent of time t. Then the expectation value has tor A, the time dependence  $ % $ % ˆ Aˆ (t) = (t)A(t) .

(7.159)

This is called the Schrödinger picture. (b) Each observable operator Aˆ changes with time t and the state vector  is independent of time. The expectation value has the time dependence $

% $  % ˆ ˆ A(t) =  A(t) .

(7.160)

This is known as the Heisenberg picture. ˆ (c) Both the state vector (t) and observable operator A(t) change with time t. This is known as the interaction picture, which is a combination of the Schrödinger and the Heisenberg pictures. To describe the Schrödinger picture, it is convenient to use the timedependent Schrödinger equation (7.93) and to rewrite it in terms of a timedependent state vector (t) and the Hamiltonian H: i


d ˆ (t) = H(t), dt

(7.161)

where the total time derivative is used here since the dependence of  on the space coordinates or other variables do not appear explicitly. The solution of

396

Chapter 7

Quantum Mechanics

(7.161) is given by ,  i t ˆ ˆ Hdt 0 = U(t) (t) = exp − 0,
0

(7.162)

ˆ is where 0 = (0), or more explicitly, 0 = (r, 0), and the operator U(t) given by ,     i t ˆ ˆ ˆ , U(t) = exp − Hdt = exp if H (7.163)
0 and   1 ˆ =− f H




t

ˆ Hdt

(7.164)

0

ˆ Since the adjoint of U(t) ˆ is a real function of the Hamiltonian operator H. is    ˆ , Uˆ ∗ (t) = exp −if H (7.165a) then Uˆ ∗ Uˆ = Uˆ Uˆ ∗ = I .

(7.165b)

ˆ Clearly, U(t) is a unitary operator. Consequently, for any  and , we obtain  $ % $  ∗ % ˆ U ˆ ˆ U = Uˆ U = |. (7.166) ˆ does not depend on time t explicitly (dH/dt ˆ In particular, if H = 0), then the solution (7.162) becomes   ˆ iH ˆ (7.167) (t) = exp − t (0) = U(t)(0),
ˆ where the unitary operator U(t) has the form   ˆ iHt ˆ U(t) = exp − .
More generally, the solution (7.167) has the form , i ˆ t0 )(t0 ). ˆ (t) = exp − H(t − t0 ) (t0 ) = U(t,


(7.168)

(7.169)

ˆ It follows from the properties of unitary operators that U(t) has an inverse −1 −1 ∗ ˆ ˆ ˆ ˆ ˆ U (t), U = U and U0 = 0 . So U(t) is isometric. The upshot of this analysis is that if a physical system is not disturbed by ˆ an experiment, there exists a unitary operator U(t), defined on the infinite

397

7.6 The Schrödinger Picture

dimensional Hilbert space of the system, which describes the time evolution of the state vector from a given time to any time t: ˆ t0 )(t0 ), (t) = U(t,

(7.170)

where Uˆ is called the time-evolution (or time-development) operator. The opˆ t0 ) does not depend on (t0 ). The immediate consequence is that erator U(t, ˆ 2 , t1 )(t1 ) = U(t ˆ 2 , t1 )U(t ˆ 1 , t0 )(t0 ) = U(t ˆ 2 , t0 )(t0 ). (t2 ) = U(t Hence, the time-evolution development operator has the property ˆ 2 , t1 )U(t ˆ 1 , t0 ). ˆ 2 , t0 ) = U(t U(t

(7.171)

It follows from (7.170) that ˆ t) = I . U(t,

(7.172)

Thus, ˆ t0 )U(t ˆ 0 , t) = U(t ˆ 0 , t)U(t, ˆ t0 ) = I . U(t, In other words, ˆ 0 , t). Uˆ −1 (t, t0 ) = U(t

(7.173)

Writing t2 = t and t1 = t − δt where δt is infinitesimal in (7.171), we obtain ˆ t − δt)U(t ˆ − δt, t0 ). ˆ t0 ) = U(t, U(t,

(7.174)

ˆ t − δt) is an infinitesimal unitary operator, which may be written in Now, U(t, the form i ˆ t − δt) = 1 − δt H(t), ˆ U(t, (7.175)
ˆ where the factor i/
is introduced for convenience of interpretation, and H(t) is a Hermitian operator, and is usually referred to as the generator of the infinitesimal unitary transformation. We next substitute (7.175) into (7.174) to obtain i ˆ − δt, t0 ) − δt H(t) ˆ U(t ˆ − δt, t0 ) ˆ t0 ) = U(t U(t,
or  1 1 ˆ U(t, ˆ t0 ). ˆ − δt, t0 ) = H(t) ˆ t0 ) − U(t lim U(t, δt→0 δt i
This leads to the evolution equation i


∂ ˆ U(t, ˆ t0 ). ˆ t0 ) = H(t) U(t, ∂t

(7.176)

398

Chapter 7

Quantum Mechanics

Obviously, its complex conjugate equation is −i


∂ ∗ ˆ Uˆ (t, t0 ) = Uˆ ∗ (t, t0 )H(t). ∂t

(7.177)

ˆ t0 ) for any time It follows from (7.176) that this equation determines U(t, ˆ 0 , t0 ) = 1. So an application of the operator t with the initial condition U(t equation (7.176) to the state vector (t0 ) gives i


∂ ˆ t0 )(t0 ) = H(t) ˆ U(t, ˆ t0 )(t0 ) U(t, ∂t

which is, by (7.170), ∂ ˆ = H(t)(t). (7.178) ∂t This is the time-dependent Schrödinger equation for the state vector (t). The time rate of change of the expectation value of an observable Aˆ in the state (t) in the Schrödinger picture is given by i


d $ ˆ % d $  ˆ % A =  A . dt dt By direct differentiation of the right-hand side, we find  2 1  1 2  ∂ d $ˆ% ∂  ˆ  ˆ A +  A A = . dt ∂t  ∂t

(7.179)

(7.180)

ˆ by using (7.93) This can be expressed in terms of the Hamiltonian operator H and its Hermitian conjugate −i


 $ ∂ ˆ | = (t)H, ∂t

(7.181)

in the form  % i$  % d $ˆ% i $ ˆ −  Aˆ H ˆ A ˆ A = H dt

2 1  i   ˆ ˆ ˆ ˆ =   H A − AH 
1  2 i   ˆ ˆ =   H, A 
1 2 i  ˆ ˆ = H, A .


(7.182)

This is the basic time-evolution equation for the expectation value of an obˆ then [H, ˆ = 0, and ˆ commutes with A, ˆ A] servable operator. In particular, if H

7.6 The Schrödinger Picture

399

ˆ is conserved the expectation value of Aˆ is a constant of motion, that is, A 2 ˆ is a ˆ ˆ A ] = 0, then A (or invariant with respect to time t). Moreover, if [H, 2 2 2 ˆ is also a constant of time. In fact, constant of time, and Aˆ = Aˆ  − A ˆ any function of an observable A has a conserved expectation value. The observable is then said to be conserved. In particular, a particle which is initially in an eigenstate of Aˆ belonging to an eigenvalue λ remains in that condition ˆ always commutes with itself, then H ˆ and H ˆ at all times. Finally, since H are always conserved, which leads to the result in classical mechanics that the energy is a constant of the motion. If the quantum system is in one of the stationary states given by ,   i ψ = ψn = ϕn (r) exp − En t ,
then

 % $   % $  ˆ H ˆ H ˆ ψn = ϕn  A, ˆ ϕn ψn  A, $  %  ˆ n, = En − En ϕn Aϕ = 0,

ˆ n = En ϕ n since Hϕ

since En = En .

It follows from (7.182) that d $ˆ% A =0 dt for the stationary state. In the presence of an external field, Aˆ may depend on time t; then an addiˆ tional term | ∂∂tA  must be added to the right-hand side of (7.180) so that (7.182) becomes 1  ˆ  2  ∂A i  ˆ d $ˆ% ˆ A  . A =   + H, dt ∂t
The right-hand side is clearly equal to the expectation value of the operator in the brackets so that 1 ˆ 2 d $ˆ% i  ˆ ∂A ˆ A A = + H, dt ∂t
1 ˆ2 i $ ˆ % ∂A ˆ A , + H, (7.183) = ∂t
or, equivalently,

1 ˆ2 $ % d $ˆ% ∂A ˆ H ˆ . + A, i
A = i
dt ∂t

(7.184)

400

Chapter 7

Quantum Mechanics

ˆ It is then possible to define an operator dA/dt and call it the total time derivˆ ative of A, by requiring the equality of the expectation values 1 ˆ2 d $ˆ% dA A (7.185) = dt dt for every state. In view of (7.184), the dynamical equation (7.183) assumes the operator form dAˆ ∂ Aˆ 1 ˆ  ˆ . = + A, H (7.186) dt ∂t i
We shall now give several examples of concrete applications of (7.182) and (7.183): (a) Aˆ = pˆ = −i
∇. ˆ Since pˆ does not involve time explicitly, ∂ p/∂t = 0. The commutator  2   2    pˆ pˆ ˆ pˆ = H ˆ pˆ − pˆ H ˆ = H, + Vˆ pˆ − pˆ + Vˆ = Vˆ pˆ − pˆ Vˆ 2m 2m gives the result 

   ˆ pˆ ψ = −i
V∇ψ ˆ ˆ ˆ H, + i
∇ Vψ = i
ψ∇ V.

Substituting this result into (7.183), we obtain $ % d ˆ = − ∇ Vˆ = F, p dt

(7.187)

where F is the classical force field. This result is known as Ehrenfest’s theorem. For a free particle, Vˆ ≡ 0; then (7.187) becomes d ˆ = 0. p dt

(7.188)

This corresponds to the classical conservation law of linear momentum for a free particle. (b) Aˆ = rˆ . Then ∂ rˆ /∂t = 0, and 

 ˆ rˆ = Hˆ ˆ r − rˆ H ˆ = H,



  2  pˆ pˆ 2 ˆ ˆ + V rˆ − rˆ +V 2m 2m

7.7 The Heisenberg Picture

=

1  2  pˆ , rˆ , 2m

401 (7.189)

since Vˆ commutes with rˆ , both being multiplicative operators. The nth component of the commutator on the right-hand side of (7.189) is  2   2      pˆ , rˆn = pˆ n , rˆn = pˆ n , rˆn pˆ n + pˆ n pˆ n , rˆn = −2i
pˆ n and so, in general,



 ˆ pˆ 2 , rˆ = −2i
p.

(7.190)

Substituting this result into (7.189), Equation (7.183) gives d 1 ˆ ˆr = p. dt m

(7.191)

This corresponds to the classical result m

dr = p. dt

(7.192)

ˆ = Tˆ + Vˆ so that the total energy is such that the potential Vˆ does (c) Aˆ = H not depend on time t. Consequently, ˆ ∂H = 0 and ∂t Equation (7.183) implies that



 ˆ H ˆ = 0. H,

d$ % ˆ = 0. H dt

(7.193)

This result corresponds to the classical conservation law of total energy. Similarly, all the conservation laws of classical mechanics can be reproduced in this way. Finally, all three applications (a) to (c) are remarkable examples of the Correspondence Principle in the sense that the expectation values of observables in quantum mechanics behave in the same way as the observables themselves do in classical mechanics.

7.7 The Heisenberg Picture and the Heisenberg Equation of Motion We denote a state vector and an observable operator in the Schrödinger picture by S and Aˆ S , and in the Heisenberg picture by H and Aˆ H . According to the

402

Chapter 7

Quantum Mechanics

Heisenberg picture, Aˆ H depends on time t, but H does not depend on t. If we define H (t) by H (t) = Uˆ −1 (t, t0 )S (t), (7.194) ˆ t0 ) is a unitary operator, then where H (t) is independent of t, and U(t, (7.194) represents a unitary operator in the vector space. In view of (7.170), the result becomes ˆ t0 )S (t0 ) = S (t0 ). H (t) = Uˆ −1 (t, t0 )U(t,

(7.195)

So it turns out that a Heisenberg operator Aˆ H is related to a Schrödinger operator Aˆ S by the unitary transformation law ˆ t0 ) Aˆ H = Uˆ −1 (t, t0 )Aˆ S U(t, ˆ t0 ). = Uˆ ∗ (t, t0 )Aˆ S U(t,

(7.196)

Differentiating this result with respect to t gives dAˆ H ∂ Uˆ ∗ ˆ ∂ Uˆ = . AS Uˆ + Uˆ ∗ Aˆ S dt ∂t ∂t

(7.197)

In view of the differential equations (7.176) and (7.177), Equation (7.197) takes the form  dAˆ H 1  ∗ˆ ˆ S Uˆ − Uˆ ∗ H ˆ S Aˆ S Uˆ = Uˆ AS H dt i
 1  ∗ˆ ˆ S Uˆ − Uˆ ∗ H ˆ S Uˆ Uˆ ∗ Aˆ S Uˆ = Uˆ AS Uˆ Uˆ ∗ H i
 1 ˆ ˆH −H ˆ H Aˆ H = AH H i
 1 ˆ ˆH . AH , H = i


(7.198)

If Aˆ H depends on time t explicitly, then (7.198) can be generalized to  dAˆ H ∂ Aˆ H 1 ˆ ˆH , = + AH , H dt ∂t i


(7.199)

where the first term on the right-hand side of (7.199) is defined by ∂ Aˆ S ∗ ∂ Aˆ H = Uˆ Uˆ . ∂t ∂t

(7.200)

Equations (7.198) and (7.199) are the celebrated Heisenberg equations of motion for the operator Aˆ H . They are similar in form to the Hamilton equations

7.7 The Heisenberg Picture

403

of motion in classical mechanics. The strong resemblance between (7.199) and (7.36) suggests that quantum analogues of the classical equations of motion can be obtained in general by substituting the commutator bracket divided by i
for the Poisson bracket {A, B} →

1 ˆ  A, Bˆ . i


(7.201)

Note that the Schrödinger Equation (7.93) is most suitable for quantum mechanic calculations. The Heisenberg equation (7.198) or (7.199) is more closely related to classical theory and can be shown to imply that classical mechanics is indeed the limit of quantum mechanics as
→ 0. For the basic canonical operators qˆ j and pˆ j , Equation (7.198) assumes the form   ˆ ˆ  1 dˆqj 1 ∂H ∂H ˆ = i
, = qˆ j , H = dt i
i
∂ pˆ j ∂ pˆ j   ˆ ˆ  1 dpˆ j ∂H 1 ∂H ˆ = . =− = pˆ j , H −i
dt i
i
∂ qˆ j ∂ qˆ j These are identical with the Hamilton equations of motion in classical mechanics.

Theorem 7.7.1. The equations of motion for the expectation values in the Heisenberg picture and in the Schrödinger picture are the same. Proof: We have % d $ˆ % d $  ˆ AH = H AH (t)H dt dt 2 1  ˆ  dA  = H  H dt 1 2  1 ˆ ˆH , = AH , H i


(7.202)

where Aˆ H does not depend on time explicitly, and H is independent of time. However, if Aˆ H depends on time t explicitly, then 1 ˆ 2 1 2  1 ˆ d $ˆ % ∂ AH ˆ + AH = AH , HH . dt ∂t i


(7.203)

Thus, the equations of motion (7.202) and (7.203) for Aˆ H  are identically the same for Aˆ S , which satisfies (7.182) and (7.183).


404

Chapter 7

Quantum Mechanics

It follows that the Heisenberg picture is more akin to classical dynamics than the Schrödinger picture. The latter does, in fact, emphasize the ˆ Hamilton–Jacobi formalism of classical mechanics, and the operator dA/dt in the Schrödinger picture is defined by (7.186). In view of the fact that the Heisenberg Equation (7.198) refers to operators, it is not of much practical importance in particular quantum mechanic problems. This leads to a dependence on all the expectation values of the operˆ commutes ator for any given state . However, if any observable operator C(t) ˆ such as Lˆ 2 or m with H, ˆ z , then the following conclusions are evident. It follows from (7.198) that  1 ˆ dCˆ ˆ = 0. = C(t), H dt i


(7.204)

Taking the expectation value for any state, it follows from (7.202) that d $ˆ % C(t) = 0. dt

(7.205)

ˆ This shows that C(t) does not change with time. If the system is in an eigenstate of Cˆ at t = 0, the state will remain in the eigenstate at any subsequent ˆ time because the operator does not change with time. All such operators C(t) satisfying the commutator relation 

 ˆ ˆ =0 C(t), H

(7.206)

are also called constants of the motion. They are the generalizations of the conserved quantities of classical mechanics. For example, the total momentum operator for a free particle is a constant of motion. For a particle in any central potential field, the total angular momentum and each separate component ˆ and, hence, according to the preceding all commute with the Hamiltonian H argument, they are also constants of the motion.

Example 7.7.2. We apply the Heisenberg equation to Aˆ H (t) = xˆ (t) for the one-dimensional motion of a particle moving in a potential V(ˆx) so that the ˆ is given by Hamiltonian operator H ˆ = H

pˆ 2 + V(ˆx). 2m

Then i


 dˆx(t)  ˆ = xˆ (t), H(t) dt ˆ −H ˆ xˆ = xˆ H

7.8 The Interaction Picture   2  pˆ pˆ 2 + V(ˆx) − + V(ˆx) xˆ = xˆ 2m 2m  1  2 = xˆ pˆ − pˆ 2 xˆ 2m  1  = [ˆx, pˆ ]pˆ + pˆ [ˆx, pˆ ] 2m 2i
pˆ i
pˆ = . = 2m m

405



Therefore, dˆx(t) pˆ (t) = . dt m

(7.207)

ˆ = pˆ (t), then Similarly, if A(t) i


 dpˆ (t)  ˆ = pˆ (t), H dt   2   2 pˆ pˆ +V − + V pˆ (t) = pˆ (t) 2m 2m   ∂ ∂ = −i
V + i
V 1 ∂x ∂x = −i


∂V(x) . ∂x

Hence, d ∂V(x) . (7.208) pˆ (t) = − dt ∂x Equations (7.207) and (7.208) may readily be generalized to three dimensions. Equation (7.208) is the direct generalization to operators of Newton’s second law of motion. This shows that the Schrödinger Equation (7.93) or equivalently the Heisenberg Equation (7.198) implies that the time-dependent observable-operators, defined by (7.196), satisfy exactly the same equations as the corresponding classical variables.


7.8 The Interaction Picture ˆ In this picture, the total Hamiltonian operator H(t) is expressed as the sum of two terms ˆ ˆ (0) + H ˆ (1) (t), H(t) =H

(7.209)

406

Chapter 7

Quantum Mechanics

ˆ (0) is the time-independent term representing the Hamiltonian of the where H ˆ (1) (t) represents the timesystem in the absence of an external field, and H dependent term that arises in the presence of an external field. In the absence of the latter term, the time-evolution operator is obtained from (7.168) in the form , i ˆ (0) . Uˆ 0 (t, t0 ) = exp − (t − t0 )H (7.210)
Both the state vector I (t) and the operator Aˆ I (t) depend on time t and are defined by I (t) = Uˆ 0−1 (t, t0 )S (t) = exp

,

i (0) ˆ (t − t0 )H S (t),


Aˆ I (t) = Uˆ 0−1 (t, t0 )Aˆ S Uˆ 0 (t, t0 ) , , i i (0) ˆ (0) ˆ ˆ AS exp − (t − t0 )H , = exp (t − t0 )H



(7.211)

(7.212)

where S (t) is the state vector and Aˆ S is the operator in the Schrödinger picture so that i


 (0)  ∂S ˆ ˆ +H ˆ (1) (t) S (t), = H(t) S (t) = H ∂t ∂ Aˆ S dAˆ S = . dt ∂t

(7.213) (7.214)

It follows from (7.211) and (7.213) that i


 , ∂ ∂I i ˆ (0) S (t) = i
exp (t − t0 )H ∂t ∂t
, -  ∂ i ˆ (0) I + exp (t − t0 )H ˆ (0) i
S = −H
∂t ,   (0) i ˆ (0) H ˆ (1) (t) S (t) ˆ +H ˆ (0) I + exp (t − t0 )H = −H
ˆ (0) I + Uˆ 0−1 (t, t0 )H ˆ (1) (t)Uˆ 0 (t, t0 )I (t) ˆ (0) I + H = −H ˆ I(1) (t)I (t), =H

(7.215)

where ˆ (1) Uˆ 0 (t, t0 ). ˆ I(1) (t) = Uˆ 0−1 (t, t0 )H H

(7.216)

7.9 The Linear Harmonic Oscillator

407

On the other hand, it also follows from (7.212) and (7.214) that  dAˆ I ∂ Aˆ I 1 ˆ ˆ I(0) , AI , H = + dt ∂t i


(7.217)

ˆ (0) Uˆ 0 (t, t0 ) = H ˆ (0) . ˆ I(0) = Uˆ 0−1 (t, t0 )H H

(7.218)

where These results show that the state vector I (t) in the interaction picture satisˆ I(1) , while the opfies the Schrödinger equation (7.215) with the Hamiltonian H erator Aˆ I (t) obeys the Heisenberg equation with the time-independent Hamilˆ (0) . tonian H

7.9 The Linear Harmonic Oscillator According to classical mechanics, a harmonic oscillator is a particle of mass m moving under the action of a force F = −mω2 x. The equation of motion is then d2 x + ω2 x = 0. (7.219) dt 2 The solution of this equation with the initial conditions, x(0) = a, x˙ (0) = 0 is x = a cos ωt.

(7.220)

This represents an oscillatory motion of angular frequency ω and amplitude a. The potential is related to the force by F = −∂V/∂x so that V(x) = 12 mω2 x2 . The energy of the oscillatory motion is the potential energy when the particle is at the extreme position. Therefore, the total energy is 1 E = ma2 ω2 . 2

(7.221)

Since the amplitude a can have any non-negative value, the total energy E can have any value greater than or equal to zero. In other words, the energy forms a continuous spectrum. We next consider the quantum theory of the harmonic oscillator. The total energy of the system is represented by the Hamiltonian operator pˆ 2 1 ˆ = Tˆ + Vˆ = H + mω2 xˆ 2 . 2m 2

(7.222)

It is convenient to introduce two dimensionless operators aˆ and aˆ ∗ by . i m aˆ = ωˆx + √ pˆ , (7.223) 2 2m

408

Chapter 7

Quantum Mechanics

. ∗

aˆ =

m i ωˆx − √ pˆ . 2 2m

(7.224)

Since xˆ and pˆ are Hermitian operators, it follows that ψ1 |ˆaψ2  = ˆa∗ ψ1 |ψ2  and

ψ1 |ˆa∗ ψ2  = ˆaψ1 |ψ2 

for any two wave functions ψ1 and ψ2 . Thus, the operators aˆ and aˆ ∗ are not Hermitian and hence they do not represent physical observables. However, aˆ aˆ ∗ and aˆ ∗ aˆ are Hermitian operators, because they can be represented as real ˆ functions of H: pˆ 2 mω2 + xˆ − 2m 2 mω2 pˆ 2 + xˆ + aˆ ∗ aˆ = 2m 2

aˆ aˆ ∗ =

iω ˆ + [ˆx, pˆ ] = H 2 iω ˆ − [ˆx, pˆ ] = H 2

1
ω, 2 1
ω, 2

ˆ can be written in terms of aˆ and aˆ ∗ as and hence H 1 1 ˆ = aˆ ∗ aˆ +
ω = aˆ aˆ ∗ −
ω H 2 2 so that [ˆa, aˆ ∗ ] =
ω,



 ˆ aˆ ∗ = [ˆa, aˆ ∗ ]ˆa∗ =
ωˆa∗ . H,

(7.225)

(7.226)

The eigenstate of energy En is |En  and ˆ n  = En |En . H|E Using (7.225), we rewrite (7.227) either as   1 ∗ aˆ aˆ |En  = En −
ω |En  2 or

  1 aˆ aˆ ∗ |En  = En +
ω |En . 2

Multiplying (7.228) by aˆ , we obtain   1 ∗ aˆ aˆ aˆ |En  = En −
ω aˆ |En . 2

(7.227)

(7.228)

(7.229)

(7.230)

Then, either aˆ |En  = 0

(7.231)

aˆ |En  = |En−1 .

(7.232)

or

7.9 The Linear Harmonic Oscillator

This result is used to rewrite (7.230) as   1 aˆ aˆ ∗ |En−1  = En −
ω +
ω |En−1 . 2

409

(7.233)

This is identical with (7.229) for En−1 , provided En−1 = En −
ω.

(7.234)

Thus, given any eigenvector |En , it is possible to generate a new eigenvector |En−1 , by (7.232), unless |En  is the lowest state |E0 . In this case, (7.231) is satisfied. It follows from (7.228) for n = 0 that 1 E0 −
ω = 0. 2

(7.235)

This determines the lowest (or ground) state energy. Clearly, it follows from (7.232) that aˆ is the operator which annihilates energy in the system in quantum units of
ω, and aˆ is called the annihilation operator. Similarly, multiplication of (7.229) by aˆ ∗ gives   1 aˆ ∗ aˆ aˆ ∗ |En  = En +
ω aˆ ∗ |En . (7.236) 2 Then, either aˆ ∗ |En  = 0

(7.237)

aˆ ∗ |En  = |En+1 .

(7.238)

or This result is used to rewrite (7.236) as   1 ∗ aˆ aˆ |En+1  = En +
ω −
ω |En+1 . 2

(7.239)

This is identical with equation (7.228) for En+1 provided En+1 = En +
ω.

(7.240)

It follows that, given any eigenstate |En , it is also possible to generate a new eigenvector |En+1 , by (7.238), with the eigenvalues given by (7.240), unless En is the highest energy level, in which case (7.237) is satisfied. But the potential is an increasing function of x and hence there is no highest level and the creation of higher energy levels is always possible. Thus, the operator aˆ ∗ generates energy in the system in quantum units of
ω and is called the creation operator. It then follows from (7.235) and (7.240) that the general energy level is   1
ω, n = 0, 1, 2, . . . . (7.241) En = n + 2

410

Chapter 7

Quantum Mechanics

This obviously represents a discrete set of energies. Thus, in quantum mechanics, a stationary state of the harmonic oscillator can assume only one of the values from the set En . The energy is thus quantized and forms a discrete spectrum. According to classical mechanics, the energy forms a continuous spectrum, that is, all non-negative numbers are allowed for the energy of a simple harmonic oscillator. This shows a remarkable contrast between the results of the classical and quantum theory. The non-negative integer n, which characterizes the energy eigenvalues (and hence eigenfunctions), is called the quantum number. The value of n = 0 corresponds to the minimum value of the quantum number with energy 1 E0 =
ω. 2

(7.242)

This is called the lowest (or ground) state energy, which never vanishes, as the lowest possible classical energy would. The ground state energy E0 is proportional to
, representing a quantum phenomenon. The discrete energy spectrum is in perfect agreement with the quantization rules of the quantum theory. To determine the energy eigenfunctions ψn belonging to En , it √ is conveˆ nient to√write the annihilation and creation operators as A ≡ aˆ /
ω and Aˆ ∗ ≡ aˆ ∗
ω and replace pˆ by −i
(∂/∂x) so that  ,  1 ∂ 1 ∂ Aˆ = √ (
mω)1/2 + (mω/
)1/2 xˆ = √ + ηˆ , (7.243) ∂x 2 2 ∂η ,   ˆA∗ = √1 −(
mω)1/2 ∂ + (mω/
)1/2 xˆ = √1 − ∂ + ηˆ , (7.244) ∂x ∂η 2 2 where ηˆ = (mω/
)1/2 xˆ . Consequently, ˆ H + Aˆ Aˆ ∗ =
ω ˆ H − Aˆ ∗ Aˆ =
ω

1 , 2

(7.245a)

1 . 2

(7.245b)

ˆ 0= Since ψ0 is the eigenfunction corresponding to the lowest energy, E0 , Aψ 0 or dψ0 + ηψ0 = 0. (7.246) dη Its normalized solution can be written as   mω 1/4 −η2 /2 e . ψ0 (η) = π


(7.247)

7.9 The Linear Harmonic Oscillator

411

All other eigenfunctions ψn can be calculated from ψ0 by successive applications of the creation operator Aˆ ∗ , and thus ψn is proportional to (Aˆ ∗ )n ψ0 . We also note that  2 1  ˆ  H $ ∗  ∗ % $  ∗ % 1 Aˆ ψn Aˆ ψn = ψn Aˆ Aˆ ψn = ψn  + ψn = (n + 1)ψn |ψn 
ω 2 (7.248) so that if ψn is normalized, so is ψn+1 = (n + 1)−1/2 Aˆ ∗ ψn . Thus, it turns out that  n ψn = (n!)−1/2 Aˆ ∗ ψ0 n  2   −1/2 η d . (7.249) = 2n n! − + η exp − dη 2 This result can be simplified by using the operator identities   d d 2 2 − + η = −eη /2 e−η /2 , dη dη n  dn d 2 2 − + η = (−1)n eη /2 n e−η /2 , dη dη

(7.250a) (7.250b)

so that the final form of ψn is   , n  n −1/2 mω 1/4 −η2 /2 n η2 d −η2 ψn (η) = 2 n! e e (−1) e (7.251) π
dηn   −1/2 mω 1/4 −η2 /2  e Hn (η), n = 0, 1, 2, . . . , (7.252) = 2n n! π
where the result in the square brackets in (7.251) defines Hn (η), the Hermite polynomials of degree n.

Example 7.9.1. (The Schrödinger equation treatment of Planck’s simple harmonic oscillator) The quantum mechanical motion of the Planck oscillator is described by the one-dimensional Schrödinger equation   d2 ψ 2M 1 2 2 + 2 E − Mω x ψ = 0. (7.253) dx2
2 In terms of the constants 2ME ,
2 Mω > 0, α=


β=

(7.254a) (7.254b)

412

Chapter 7

Quantum Mechanics

√ and an independent variable x = x α, Equation (7.253) becomes, dropping the prime,   d2 ψ β − x ψ = 0. (7.255) + dx2 α The eigenfunctions of this equation are the Hermite orthogonal functions    √  αx2 (7.256) ψn (x) = An Hn x α exp − 2 with the corresponding eigenvalues β = (2n + 1), α

(7.257)

where Hn (x) is the Hermite polynomial of degree n. Substituting the values of α and β, it turns out that   2n + 1 E ≡ En = ω
, n = 0, 1, 2, . . . . (7.258) 2 The so-called half-integral multiples of the energy quanta are the characteristics of the oscillator, that is, the odd multiples of 12 ω
. This result is remarkably the same as in the Heisenberg theory. In view of the following properties of the Hermite polynomials H0 (x) = 1,

H1 (x) = 2x,

and

H2 (x) = 4x2 − 2,

it follows that the first eigenfunction ψ0 (x) represents a Gaussian distribution curve, and the second eigenfunction ψ1 (x) vanishes at the origin and corresponds to a Maxwellian distribution curve for positive x, which is continued toward negative values of x so that it is an odd function of x. The third eigen√ function ψ2 (x) is negative at the origin and has two symmetric zeros ±1/ 2 and so on. Thus, the geometrical shape of these eigenfunctions can easily be determined. It is also important that the roots of successive polynomials separate one another.


7.10 Angular Momentum Operators The orbital angular momentum operators Lˆ x , Lˆ y , and Lˆ z were introduced in Section 7.3. It has been shown that they obey the commutation relations (7.66). Using the spherical polar coordinates (r, θ, ϕ), which are related to rectangular Cartesian coordinates (x, y, z) by x = r sin θ cos ϕ,

(7.259a)

7.10 Angular Momentum Operators

413

y = r sin θ sin ϕ,

(7.259b)

z = r cos θ,

(7.259c)

combined with the chain rule for differentiation ∂ ∂r ∂ ∂θ ∂ ∂ϕ ∂ = + + ∂x ∂x ∂r ∂x ∂θ ∂x ∂ϕ and similar results for ∂/∂y and ∂/∂z, the angular momentum operators can be expressed in angular variables   ∂ ∂ + cot θ cos ϕ Lˆ x = i
sin ϕ , (7.260a) ∂θ ∂ϕ   ∂ ∂ ˆLy = i
− cos ϕ + cot θ sin ϕ , (7.260b) ∂θ ∂ϕ Lˆ z = −i


ˆ2

L

= Lˆ 2x

+ Lˆ 2y

+ Lˆ 2z

∂ , ∂ϕ ,

= −


2

(7.260c)

  ∂ 1 ∂2 1 ∂ sin θ + 2 . sin θ ∂θ ∂θ sin θ ∂ϕ 2

From (7.259c) and (7.261) it is easy to verify that  2  Lˆ , Lˆ z = 0.

(7.261)

(7.262)

It also follows from (7.259c) and (7.157) that Lˆ z Ylm (θ, ϕ) = (
m)Ylm (θ, ϕ).

(7.263)

For any given value of l, the possible eigenvalues of the z-component of the angular momentum, Lˆ z are Lz ≡
m,

m = 0, ±1, ±2, . . . , ±l,

(7.264)

giving (2l + 1) admissible values. On the other hand, it can easily be verified with the aid of (7.261), (7.147), and (7.157) that, with λ = l(l + 1),   Lˆ 2 Ylm (θ, ϕ) =
2 l(l + 1) Ylm (θ, ϕ), (7.265) where |m| ≤ l, and l = 0, 1, 2, . . . . This shows that the eigenvalues of Lˆ 2 are L2 ≡
2 l(l + 1),

l = 0, 1, 2, 3, . . . .

(7.266)

414

Chapter 7

Quantum Mechanics

Evidently, the spherical harmonics Ylm (θ, ϕ) are the simultaneous eigenfunctions of Lˆ 2 and Lˆ z . The eigenvalues of the total angular momentum Lˆ 2 are
2 l(l + 1), l = 0, 1, 2, . . . , and those of Lˆ z are m
, m = 0, ±1, . . . , ±l. Thus, a measurement of L2 can yield as its result only the values 0, 2
2 , 6
2 , 12
2 , . . . . The total angular momentum states with l values 0, 1, 2, 3, 4 are known, for historical reasons, as S, P, D, F, G states, respectively. Similarly, the measured values of Lˆ z are only 0, ±
, ±2
, . . . . Hence, both Lˆ 2 and Lˆ z are quantized and can only reveal one of the specified discrete values upon measurement. It is convenient to define two operators Lˆ + and Lˆ − by Lˆ + = Lˆ x + iLˆ y ,

(7.267a)

Lˆ − = Lˆ x − iLˆ y .

(7.267b)

Theorem 7.10.1. (a) Lˆ + and Lˆ − are non-Hermitian operators, (b) Lˆ + Lˆ − and Lˆ − Lˆ + are Hermitian.

Proof: Since Lˆ x and Lˆ y are Hermitian, $ % $ % Lˆ + ψ1 |ψ2 = ψ1 |Lˆ − ψ2 ,

$ % $ % Lˆ − ψ1 |ψ2 = ψ1 |Lˆ + ψ2

for any two wave functions ψ1 and ψ2 . Thus, Lˆ + and Lˆ − are not Hermitian operators, and hence they do not represent observables. On the other hand,      Lˆ + Lˆ − = Lˆ x + iLˆ y Lˆ x − iLˆ y = Lˆ 2x + Lˆ 2y − i Lˆ x , Lˆ y   = Lˆ 2x + Lˆ 2y +
Lˆ z = Lˆ 2 − Lˆ z Lˆ z −
. (7.268) Similarly,

  Lˆ − Lˆ + = Lˆ 2x + Lˆ 2y −
Lˆ z = Lˆ 2 − Lˆ z Lˆ z +
.

(7.269)

Thus, both Lˆ + Lˆ − and Lˆ − Lˆ + are expressed as real functions of Lˆ 2 and Lˆ z . Hence, they are Hermitian operators.
Since the orbital angular momentum can only take on integer values, this result indicates the necessity for some generalization of this formalism. It is necessary to introduce matrix operators of size n × n defined by  1|A1 ˆ ˆ  2|A1 ˆA =  .  .. ˆ n|A1

ˆ 1|A2 ˆ 2|A2 .. .

ˆ n|A2

... ... .. .

ˆ  1|An ˆ  2|An , ..  . 

ˆ . . . n|An

ˆ , i = 1, 2, . . . , n, j = 1, 2, 3, . . . , n are the elements of the matrix. where i|Aj

7.10 Angular Momentum Operators

415

An eigenvalue for the matrix operator, Aˆ is defined by ˆ n  = an |un , A|u

(7.270)

ˆ respectively. where an and |un  are the eigenvalues and eigenvectors of A, If the matrix Aˆ is of the diagonal form   ˆA = a1 0 , (7.271) 0 a2 it is very easy to check by direct substitution into (7.270) that the eigenvalues are a1 and a2 with eigenvectors   1 , (7.272a) |u1  = 0   0 |u2  = . (7.272b) 1 We next assume the definition of the orbital angular momentum operators given in Section 7.3 and derive the commutation relations (7.66). We then take these commutation relations as the definition of the general angular momenˆ = (M ˆ x, M ˆ y, M ˆ z ). tum operators M

Definition 7.10.2.

(The Pauli spin matrices) The Pauli spin matrices are defined by      0 1 0 −i 1 σˆ x = , σˆ y = , σˆ z = 1 0 i 0 0

 0 . −1

(7.273)

The Pauli spin matrices were defined by Wolfgang Ernst Pauli (1900–1958) in 1927. These matrices satisfy the following relations σˆ x σˆ z = −σˆ y σˆ x = iσˆ z ,

(7.274a)

σˆ y σˆ z = −σˆ z σˆ y = iσˆ x ,

(7.274b)

σˆ z σˆ x = −σˆ x σˆ z = iσˆ y .

(7.274c)

It is very simple to check that the matrices 1 ˆ x =
σˆ x , M 2 1 ˆ y =
σˆ y , M 2

(7.275a) (7.275b)

416

Chapter 7

Quantum Mechanics

1 ˆ z =
σˆ z M 2 satisfy the commutation relations (7.66). Also it follows directly that    1 1 1 2 2 2 2 ˆx +M ˆy +M ˆz = ˆ =M + 1
2 M 0 2 2 so that



(7.275c)

 0 , 1

  2   2  ˆ 2, M ˆx = M ˆy = M ˆ z = 0. ˆ ,M ˆ ,M M

(7.276)

(7.277)

ˆ z are According to (7.271), (7.272a), and (7.272b), the eigenvalues of M 1 Mz = ±
2

(7.278)

with the corresponding eigenvectors   1 , |u+1/2  = 0   0 |u−1/2  = . 1

(7.279a) (7.279b)

ˆ 2 , each corresponding to the eigenThe same vectors are also eigenvectors of M value   1 1 2 M = (7.280) + 1
2 . 2 2 The results (7.278) and (7.280) are precisely the same as (7.264) and (7.266), which were established for the orbital angular momentum, except that in orbital angular momentum operators the possible values of l and m were restricted to integers. With the commutation relations (7.66) as the defining ˆ x, M ˆ y, M ˆ z , and allowing maproperty for the angular momentum operators M trix representation of these operators, it clearly indicates the possibility of nonintegral values for l = 12 and m = ± 12 . Since this is not associated with orbital motion, it must be the intrinsic angular momentum or spin of the particle itself. The so-called electron spin thus represents an example of an angular momentum 12 . Physically, an electron as an electrically charged particle interacts with an external magnetic field. The effect of the interaction is a shifting in atomic energy levels which results in a change in spectral lines. The observed spectrum of an atom in the presence of a magnetic field is known as the anomalous Zeeman effect. Experimental facts in spectroscopy including the Zeeman effect led to the discovery of the electron spin. Many other spectroscopic phenomena such as the multiplicity of numerous lines cannot be explained without

7.10 Angular Momentum Operators

417

the introduction of this new hypothesis. So the new hypothesis has not only resolved many important difficulties, but is essential for an understanding of many atomic and nuclear phenomena. To establish the possibility of both integer and half-integer values of the angular momentum, the operator techniques can be applied to the problem of the angular momentum. We take the commutation relations as the defining property of the angular momentum operators M:   ˆ x, M ˆ y = i
M ˆ z, M (7.281a)   ˆ z = i
M ˆ x, ˆ y, M (7.281b) M   ˆ x = i
M ˆ y. ˆ z, M (7.281c) M It is convenient to introduce two new operators ˆ x + iM ˆ y, ˆ+≡M M

(7.282a)

ˆ x − iM ˆ y. ˆ−≡M M

(7.282b)

By direct substitution, it turns out from (7.281a) to (7.281c) that   ˆ + =
M ˆ +, ˆ z, M M   ˆ − = −
M ˆ z, M ˆ −, M   ˆ − = 2
M ˆ z. ˆ +, M M

(7.283a) (7.283b) (7.283c)

ˆ z commute, we can introduce a state |λ, m, ˆ 2 and M In view of the fact that M which is simultaneously an eigenstate of both operators. Hence, ˆ 2 |λ, m =
2 λ|λ, m, M

(7.284)

ˆ z |λ, m =
m|λ, m. M

(7.285)

So the main problem is to determine the possible values of λ and m implied by the relations (7.281a) to (7.281c). For a given value of the magnitude of the ˆ z, angular momentum λ, from a physical point of view, the possible values of M determined by m, must lie in the finite domain with values bounded by mmax and mmin . This observation is important in the subsequent discussion. It now follows from (7.285) that ˆ +M ˆ z |λ, m =
mM ˆ + |λ, m. M

(7.286)

In view of (7.283a), we obtain   ˆ +M ˆz =M ˆ+− M ˆ + −
M+ , ˆ zM ˆ z, M ˆ+ =M ˆ zM M

(7.287)

418

Chapter 7

Quantum Mechanics

which is substituted into (7.286) to obtain ˆ zM ˆ + |λ, m =
(m + 1)M ˆ + |λ, m. M

(7.288)

ˆ + |λ, m = 0 M

(7.289)

ˆ + |λ, m = |λ, m + 1. M

(7.290)

Clearly, either or Consequently, (7.288) becomes ˆ z |λ, m + 1 =
(m + 1)|λ, m + 1. M

(7.291)

Thus, it follows from (7.285) and (7.291) that given an eigenstate |λ, m, a new eigenstate |λ, m + 1 can be generated by (7.290) with eigenvalue
(m + 1), unless m = mmax , in which case (7.289) holds. The admissible values of m differ by integers. ˆ z to show that eiˆ −M An argument similar to the preceding is applied to M ther ˆ − |λ, m = 0, M

(7.292)

ˆ − |λ, m = |λ, m − 1, M

(7.293)

ˆ z |λ, m − 1 =
(m − 1)|λ, m − 1. M

(7.294)

or so that So, given any eigenstate |λ, m, a new eigenstate can be generated by (7.293) with eigenvalues
(m − 1), unless m = mmin , in which case (7.292) applies. We have    ˆ+= M ˆ x − iM ˆy M ˆy ˆ −M ˆ x + iM M   ˆ y2 + i M ˆ x, M ˆy =M ˆ z2 −
M ˆ z. ˆ 2−M ˆ x2 + M (7.295) =M ˆ −M ˆ + acts on |λ, mmax , it follows by using (7.289) that If the operator M  2  ˆ −M ˆ + |λ, mmax  = 0. ˆ z2 −
M ˆ z |λ, mmax  = M ˆ −M M (7.296) Therefore, by (7.284) and (7.285),   λ − m2max − mmax |λ, mmax  = 0, or λ = m2max + mmax .

(7.297)

7.10 Angular Momentum Operators

419

Similarly, ˆ 2−M ˆ z2 +
M ˆ z. ˆ−=M ˆ +M M

(7.298)

If this operator acts on |λ, mmin  and (7.292) is used, we obtain λ − m2min + mmin = 0.

(7.299)

If we equate the two results for λ from (7.297) and (7.299), it turns out that (mmax + mmin )(mmin − mmax − 1) = 0.

(7.300)

mmax = −mmin .

(7.301)

Thus, Therefore, the admissible values of m lie symmetrically about the origin. Since the extreme values differ by an integer, it follows that mmax − mmin = 2l,

(7.302)

where 3 1 l = 0, , 1, , . . . . 2 2 These results combined with (7.301) show that −l ≤ m ≤ l

(2l + 1 values).

(7.303)

(7.304)

Finally, it follows from (7.297) and (7.302) that λ = l(l + 1),

1 3 l = 0, , 1, , . . . . 2 2

(7.305)

This is a definite proof for integer and half-integer eigenvalues for the angular momentum. Particles with integral spin are called the Bosons after S. N. Bose (1894–1974), those with half-integral spins are known as Fermions after Enrico Fermi (1901–1954). The two different kinds of angular momentum operators can be combined to define the total angular momentum ˆ Jˆ = Lˆ + M

(7.306)

ˆ x , Jˆy = Lˆ y + M ˆ y , Jˆz = Lˆ z + M ˆ z. with the components Jˆx = Lˆ x + M ˆ ˆ ˆ It follows from the properties of L and M that J satisfies the usual commutation relations   Jˆx , Jˆy = i
Jˆz , (7.307a)   (7.307b) Jˆy , Jˆz = i
Jˆx ,   (7.307c) Jˆz , Jˆx = i
Jˆy ,

420 and hence,

Chapter 7

Quantum Mechanics



     Jˆx , Jˆ2 = Jˆy , Jˆ2 = Jˆz , Jˆ2 = 0,

(7.308)

Jˆ2 = Jˆx2 + Jˆy2 + Jˆz2 .

(7.309)

Jˆ2 |l, m = l(l + 1)
2 |l, m,

(7.310)

Jˆz |l, m =
m|l, m.

(7.311)

where It can readily be shown that

This means that the eigenvalues of Jˆ2 and Jˆz are l(l + 1)
2 and
m, respectively, where |m| ≤ l and the quantum numbers may be either integers or halfintegers. Finally, it follows that 2        ˆ x Lˆ x , Lˆ z + 2M ˆ y Lˆ y , Lˆ z =
Lˆ + , Lˆ − , Jˆ , Lˆ z = 2M (7.312)    2 ˆ z = −
Lˆ + , Lˆ − . (7.313) Jˆ , M

7.11 The Dirac Relativistic Wave Equation The Schrödinger wave equation, which forms the basis of quantum mechanics, does not satisfy the requirements of the special theory of relativity, namely, invariance under the Lorenz transformations. Dirac first formulated a relativistic version of the Schrödinger equation, which is universally known as the Dirac equation. His famous generalization of the Schrödinger equation to a set of four first-order partial differential equations met the requirements of relativity and incorporated spin properties of particles without abandoning the general physical aspects of the wave-mechanical description. Historically, the Dirac equation is considered as fundamental in the sense that it led to the discovery of antiparticles, to the basic understanding of electron’s magnetic moment, and to more precise prediction of the spectra in atoms. With the standard Hamiltonian form, the Dirac equation is i


∂ψ ˆ = Hψ ∂t

(7.314)

where the Dirac Hamiltonian is ˆ = cα · pˆ + βmc2 , H

(7.315)

where α = (α1 , α2 , α3 ) and β are assumed to be dimensionless constants (independent of space and time) determined by the requirement that  2 (7.316) H 2 = (pc)2 + mc2 .

7.11 The Dirac Relativistic Wave Equation

421

Keeping in mind the operator nature of α and β, we use the condition (7.316) so that    2 cα · p + βmc2 = (pc)2 + mc2 , or, preserving the order of relevant factors, we obtain 

αi2 p2i c2 +

  2 1 (αi αj + αj αi )pi pj c2 + (αi β + βαi )mc2 p2i + β 2 mc2 2 i=j

i

i

 2 = (pc)2 + mc2 . Comparing terms, it turns out that αi2 = β 2 = 1, or α12 = α22 = α32 = β 2 = 1,

(7.317)

and α1 α2 + α2 α1 = α2 α3 + α3 α2 = α3 α1 + α1 α3 = α1 β + βα1 = α2 β + βα2 = α3 β + βα3 = 0. (7.318) The four matrices mutually anticommute and their squares are unity. So the problem is to find these matrices with the properties (7.317) and (7.318). These matrices are called the Dirac matrices whose algebra is seen to be identical to that of Pauli matrices except that there are four of them. Since the Pauli matrices (along with the unit matrix) exhaust the independent 2 × 2 matrices, the four Dirac matrices cannot be represented by 2 × 2 matrices. It turns out that 3 × 3 matrices do not suffice, and the smallest matrices that do are 4 × 4. These are not uniquely defined by the commutation relations, but the conventional choice is     1 0 0 0 0 0 0 1 0  0 1 0 0 0 1 0 β = α1 =  , , 0 0 −1 0 0 1 0 0 0 0 0 −1 1 0 0 0     0 0 0 −i 0 0 1 0 0 0 i 0   0 0 0 −1  α3 =  α2 =  , . 0 −i 0 0 1 0 0 0 i 0 0 0 0 −1 0 0 These matrices can also be expressed in terms of the 2 × 2 Pauli matrices σj as

 β=

I 0

0 −I



 ,

αj =

0 σj

σj 0

 ,

j = 1, 2, 3,

(7.319)

422

Chapter 7

Quantum Mechanics

where every element of these 2 × 2 matrices is itself to be understood as 2 × 2 matrices. Since the Dirac matrices are 4 × 4, the Dirac equation makes sense if the Dirac wave function is a column matrix with four rows as   ψ1  ψ2  ψ =  . ψ3 ψ4 The Dirac equation can be put in a more appealing form by introducing a new set of 4 × 4 matrices as γ 0 = β,

γ 1 = βα1 ,

γ 2 = βα2 ,

γ 3 = βα3 ,

(7.320)

and by replacing the momentum operator p with −i
∇ in (7.314) and (7.315). It turns out that the Dirac equation assumes the form i


 ∂ψ  = −i
cαˆ · ∇ + βmc2 ψ, ∂t

(7.321)

or, equivalently, 1 0 ∂ψ  j  γ + γ · ∇ ψ + iκψ = 0, c ∂t



 mc κ= ,


where the explicit form of the Dirac γ matrices are     I 0 0 σj 0 j , γ = , j = 1, 2, 3. γ = 0 I −σj 0 These matrices satisfy the following properties  j ∗  0 ∗ = γ 0, γ = γ j for j = 1, 2, 3, γ where ∗ represents the conjugate transpose  0 2  j 2 γ = I, γ = −I,

for j = 1, 2, 3,

(7.322)

(7.323)

(7.324)

(7.325)

and γ k γ  + γ  γ k = 0 for k = , k,  = 0, 1, 2, 3.

(7.326)

Equation (7.322) is a conventional form of the Dirac equation and its solution ψ is called the spinor wave function or simply, spinor. The Dirac equation (7.322) admits plane wave solutions of the form     ψj (r, t) = aj exp i(k · r − ωt) = aj exp −i
(p · r − Et) , j = 1, 2, 3, 4, (7.327)

7.12 Exercises

423

where aj are scalars. These solutions represent the eigenfunctions of the energy and momentum operators with eigenvalues
ω and
k, respectively. Explicit solutions can be found for momentum p by choosing a sign for the energy E+ = H+ = +(c2 p2 + m2 c4 )1/2 . Then there are two sets of linearly independent solutions for aj . Similarly, there are two sets of solutions corresponding to the negative square root of energy, H− = E− = −(c2 p2 + m2 c4 )1/2 . The former solutions correspond to positive energy and the latter to negative energy. In the nonrelativistic limit, in which H+ = −H− is close to mc2 and large compared to cp. The physical distinction between the two solutions for each sign of energy can be seen by defining three new 4 × 4 spin matrices γ˜ j as   σj 0 j , j = 1, 2, 3, (7.328) γ˜ = 0 σj where σj are the 2 × 2 Pauli matrices. The Dirac equation implies the Klein–Gordon equation. Indeed, the former is a square root of the latter in the sense -2   ,  j  1 ∂2 1 0∂ 2 2 (7.329) γ + γ · ∇ + iκ ψ = 2 2 − ∇ + κ ψ c ∂t c ∂t as operators. The equations of quantum electrodynamics form a system that combines the Dirac and Maxwell equations with nonlinear coupling terms. Physically, they describe interaction between electrons governed by the Dirac equation and photons described by the Maxwell equations. The predictions of the Dirac theory are in remarkable agreement with experimental observations, and hence it is the most accurate theory in all of physical sciences.

7.12 Exercises 1. (a) Use the Lagrangian, L = 12 m(˙x2 + y˙ 2 + z˙ 2 ) − 12 k(x2 + y2 + z2 ) for the three-dimensional isotropic harmonic oscillator, and Lagrange’s equations of motion to show that the total energy is constant where k is a constant. (b) Show that the Lagrangian for the oscillator in spherical polar coordinates (r, θ, ϕ) is  k2 1  L = T − V = m r2 + r2 θ˙ 2 + r2 sin2 ϕ ϕ˙ 2 − , 2 r where k = 4π 2 mω2 . Hence, write down the Lagrange equations of motion.

424

Chapter 7

Quantum Mechanics

2. Consider a single particle of mass m moving in a plane under a conservative force with potential V(r), where r is the distance from the origin of coordinates. With r and θ as generalized coordinates describing the motion of the particle, show that the corresponding momenta are pr =

∂L = m˙r, ∂ r˙

pθ =

∂L = mr2 θ˙ , ∂ θ˙

where L = T − V = 12 m(˙r2 + r2 θ˙ 2 ) − V(r). Hence show that H=

p2 p2r + θ 2 + V(r), 2m 2mr

mr2 θ˙ = constant,

  ∂V m r¨ − rθ˙ 2 = − . ∂r

Give an interpretation of each of these results. 3. If A is a complex dynamical function of q and p, A∗ is its complex conjugate, and if the Poisson bracket {A, A∗ } = i, compute {A, AA∗ }, {A, A∗ A}, {A∗ , AA∗ }, and {A∗ , A∗ A}. 4. Find the Hamiltonian and Hamilton’s equations of motion for (a) The simple harmonic oscillator with T = 12 m˙x2 and V = 12 kx2 and (b) the planetary motions with T = 12 m(˙r2 + r2 θ˙ 2 ), and V = mµ( 12 a − 1/r). In this case, derive the differential equations for the central orbit. 5. Establish the following results for the Poisson brackets: (a) {A, B} = −{B, A}. (b) {(A + B), C} = {A, C} + {B, C}. (c) {AB, C} = {A, C}B + A{B, C}. (d) {A, α} = 0. (e) {A, {B, C}} + {B, {C, A}} + {C, {A, B}} = 0 (Jacobi’s identity), where A, B, and C are canonical functions and α is a scalar. 6. Show that the following results hold for commutators: ˆ B] ˆ ˆ = −[B, ˆ A]. (a) [A, ˆ = [A, ˆ C] ˆ + [B, ˆ ˆ C] ˆ C]. (b) [Aˆ + B, ˆ Bˆ + C] ˆ = [A, ˆ B] ˆ C]. ˆ ˆ + [A, (c) [A, ˆ = [A, ˆ C] ˆ Bˆ + A[ ˆ B, ˆ ˆ C] ˆ C]. (d) [Aˆ B, ˆ Bˆ C] ˆ = [A, ˆ B] ˆ C]. ˆ ˆ Cˆ + B[ ˆ A, (e) [A, ˆ [B, ˆ + [B, ˆ A]] ˆ + [C, ˆ [A, ˆ B]] ˆ C]] ˆ [C, ˆ = 0 (Jacobi’s identity). (f) [A,

7.12 Exercises

425

ˆ A, ˆ B] ˆ B] ˆ ˆ = A[ ˆ + [A, ˆ A, (g) [Aˆ 2 , B] ˆ α] = 0, where α is a scalar. (h) [A, 7. For the three-dimensional position and momentum operators of a particle, prove that [ˆri , pˆ j ] = i
δij , where the suffixes i, j take the values 1, 2, 3 for the x, y, z components of ˆ respectively. rˆ and p, 8. By direct evaluation for canonically conjugate variables q and p, show that (a) [p2 , q2 ] = 2
2 − 4i
pq. (b) [p, q2 ] = −2i
q. (c) [ˆx2 , pˆ 2x ] = 2
2 − 4iˆxpˆ x . (d) [pˆ x , xˆ 2 ] = −2i
ˆx. 9. If A and B are two operators which both commute with their commutaˆ B], ˆ prove that tor [A,     ˆ Bˆ , ˆ Bˆ n = nBˆ n−1 A, A,    n  ˆ Bˆ . Aˆ , Bˆ = nAˆ n−1 A, 10. Establish the following commutator relations:  2   2   2  Lˆ , Lˆ x = Lˆ , Lˆ y = Lˆ , Lˆ z = 0. 11. Show that 

 Lˆ + , Lˆ − =
Lˆ z ,   Lˆ + , Lˆ z = −
Lˆ + ,   Lˆ − , Lˆ z =
Lˆ − ,   Lˆ x , Lˆ + =
Lˆ − ,   2 Lˆ , Lˆ + = 0.

12. Prove that ˆ 2 + 2Lˆ · M ˆ = Lˆ 2 + M ˆ 2 + 2Lˆ z M ˆ z + Lˆ + Lˆ − + Lˆ − Lˆ + , Jˆ2 = Lˆ 2 + M ˆ = Jˆ2 − Lˆ 2 − M ˆ 2. 2Lˆ · M

426

Chapter 7

Quantum Mechanics

13. Show that the probability for a position measurement on the state (x, t) to yield a value somewhere between x1 and x2 is  x2  x2   (x, t)2 dx. dx = P(x1 , x2 , t) = x1

x1

Using the Schrödinger equations, derive the result d P(x1 , x2 , t) = J(x1 , t) − J(x2 , t), dt where J(x, t) =

, i
∂ ∂  − . 2m ∂x ∂x

14. Use the inner product  ϕ, ψ =

∞

ϕψdx, −∞

and the property ϕ, ψ → (0, 0) as |x| → ∞, to show that the position operator xˆ = x, the momentum operator pˆ = −i
∂/∂x, and the energy ˆ = (pˆ 2 /2m) + V(ˆ ˆ x) are Hermitian operators. operator H 15. Establish the following commutation relations for the orbital angular momentum operators:   Lˆ x , xˆ = 0,   Lˆ x , yˆ = i
ˆz,   Lˆ x , zˆ = −i
ˆy,   Lˆ x , pˆ x = 0,   Lˆ x , pˆ y = i
pˆ z ,   Lˆ x , pˆ z = −i
pˆ y . 16. Prove the Heisenberg uncertainty relation for the harmonic oscillator 1 xp ≥
. 2 ˆ B] ˆ is 17. If Aˆ and Bˆ are constants of motion, show that the commutator i[A, also a constant of motion.

7.12 Exercises

427

18. Show that, for the linear harmonic oscillator,     ˆ ˆ Aˆ = (−
ω)A, ˆ Aˆ ∗ = (
ω)Aˆ ∗ , H, H, where

√ Aˆ = aˆ /
ω

and

√ Aˆ ∗ = aˆ ∗ /
ω.

19. For the three-dimensional anisotropic Planck’s oscillator, the Hamiltonian is given by Hr =

1 2 1 p + mωr2 xr2 , 2m r 2

r = 1, 2, 3,

so that the Hamiltonian H = H1 + H2 + H3 and the total energy E = E1 + E2 + E3 , where E1 , E2 , E3 are energies of each of the independent degrees of freedom. Show that       1 1 1
ω1 + n1 +
ω2 + n3 +
ω3 . E = n1 + 2 2 2 In the case of an isotropic oscillator, ω1 = ω2 = ω3 = ω, derive the result   3 EN = N +
ω, N = (n1 + n2 + n3 ) = 0, 1, 2, 3, . . . . 2 20. Prove the compatibility theorem which states that any one of the following conditions implies the other two: (a) Aˆ and Bˆ are compatible, (b) Aˆ and Bˆ possess a common eigenbasis, (c) Aˆ and Bˆ commute, where A and B are two observables with corresponding operators Aˆ ˆ and B. 21. If the eigenvectors {ψn (x)} form an orthonormal basis in a Hilbert space, show that any state vector ψ(x) satisfies the result ψ, ψ =

∞    ψn , ψ2 . n=1

ˆ and Bˆ ≡ Bˆ − B, ˆ prove the following results: 22. If Aˆ ≡ Aˆ − A (a) Aˆ and Bˆ are Hermitian operators. ˆ B]. ˆ (b) [Aˆ , Bˆ ] = [A,

428

Chapter 7

Quantum Mechanics

ˆ 2. (c) Aˆ ψ, Aˆ ψ = (A) Use these results to establish the generalized uncertainty relation.  ˆ = ∞  ∗ (x)[A(x)]dx, ˆ 23. Using A prove that the expectation values of −∞ position and momentum in the state (x, t) are  ∞  ∞  2 ∂   ˆx = x (x, t) dx, pˆ  = −i
 ∗ (x, t)dx. ∂x −∞ −∞ Also show that  ∞  2 $ 2% x2 (x, t) dx, xˆ =

$ 2% pˆ =
2

∞



2    ∂ (x, t) dx.   ∂x

∞ −∞

24. Apply the basic commutation relations [ˆxi , pˆ j ] = i
δij and rules of commutator algebra to show that  i
  ˆ = pˆ 2x + xˆ pˆ x , Vˆ , xˆ pˆ x , H m  i
   ˆ = pˆ 2z + zˆ pˆ z , Vˆ . zˆ pˆ z , H m





 i
  ˆ = pˆ 2y + yˆ pˆ y , Vˆ , yˆ pˆ y , H m

Combine these results to obtain the Heisenberg equation of motion for the operator r · p 1 22 p d r · p = − r · ∇V. dt m Prove the Virial Theorem for the stationary states: 2T = r · ∇V. 25. Use the results in Exercise 9 for Aˆ = xˆ and Bˆ = pˆ x to prove that for any Hamiltonian of the form ˆ = H

pˆ 2x + α xˆ n , 2m

the following relation holds:  2      pˆ ˆ = i
x − αnˆxn = i
2Tˆ − nVˆ . xˆ pˆ x , H m 26. Use the Hamiltonian operator for the one-dimensional simple harmonic oscillator in the form ˆ = H

1 2 1 pˆ + mω2 xˆ 2 , 2m 2

7.12 Exercises

429

and then introduce the nondimensional variables   mω 1/2 1 Xˆ = xˆ , Pˆ = pˆ . 2
(2m
ω)1/2 (a) Show that (i) Xˆ and Pˆ are Hermitian operators. ˆ =
ω(Pˆ 2 + Xˆ 2 ). (ii) H ˆ P] ˆ = 1 i. (iii) [X, 2 ˆ = Xˆ + iPˆ and Q ˆ ∗ = Xˆ − iP, ˆ show that (b) If Q ˆQ ˆ ∗ = Xˆ 2 + Pˆ 2 + 1 , Q 2   1 ∗ ˆ Q ˆ+ ˆ =
ω Q H , 2

ˆ ∗Q ˆ = Xˆ 2 + Pˆ 2 − 1 , Q 2

ˆ and Q ˆ ∗ is defined by the commutawhere the algebra of the operators Q tion relation   ˆ Q ˆ ∗ = 1. Q, 27. If Aˆ and Bˆ are two vector operators that commute with the Pauli spin matrices but do not commute between themselves, prove the Dirac identity        ˆ σˆ · Aˆ σˆ · Bˆ = Aˆ · Bˆ + i Aˆ × Bˆ · σ, where σˆ = (σˆ x , σˆ y , σˆ z ). 28. For the linear harmonic oscillator, establish that   ˆ aˆ = −
ωˆa. H, ˆ with eigenvalues Show that this implies that aˆ |ψ is the eigenstate of H (E −
ω). Hence, deduce that (a) E ≥ 12
ω, and (b) E = 12
ω if and only if a|ψ = 0. 29. For thefree Dirac particles, show that the Dirac Hamiltonian operator ˆ 2 does not commute with angular momentum opˆ = c k αˆ k pˆ k + βmc H erators Lˆ k where k = 1, 2, 3. 30. Show that the plane wave solutions   ψj (r, t) = aj exp −i
(p · r − Et) ,

j = 1, 2, 3, 4

430

Chapter 7

Quantum Mechanics

of the Dirac equation (7.321) lead to the relativistic energy relations  1/2 . E± = ± c2 p2 + m2 c4 31. Consider a one-dimensional simple harmonic oscillator system consisting of a single particle of mass m and a spring whose constant is k. With the canonical coordinates (q, p), p = ∂S ∂q , the Hamiltonian of this conservative system is H=

kq2 p2 + = total energy = E. 2m 2

Show that . q= where ω =

2E cos ω(t − τ ), k

√ p = − 2Em sin ω(t − τ ),

k/m is the natural frequency of this mass-spring system.

32. (a) Consider an n-dimensional conservative system with H = E so that the Hamilton principle function S and the Hamilton characteristic function W are related by S(q, p, t) = W(q, p) − Et, where q and p are n-dimensional vectors. Show that the phase velocity cp of the wave is given by cp =

E ds = , dt |∇W|

where ds is the element of the arc length normal to the surface of constant S or W. (b) For a single particle without constraints, show that the phase velocity is cp = E/p and the momentum is p = ∇W. 33. For an infinite square well with the potential V(x) = V0 H(|x| − a), the solution of the one-dimensional time-independent Schrödinger equation is

(a) ψ1 (x) = Aeκx (x ≤ −a), κ = 2m (V0 > E),
2 (V0 − E) (b) ψ2 (x) = 0, x ≥ a, (c) ψ3 (x) = A sin kx + B cos kx, for |x| ≤ a, where A, B are arbitrary constants and k2 = 2mE
2 .

7.12 Exercises

431

Hence, find the energy levels for an even solution (cos ak = 0) and for an odd solution (sin ak = 0). 34. (a) For the nonrelativistic Dirac equation, if a particle moves along the z-axis, show that there are four degenerate eigenvalues E+ , E+ , E− , and E− where E± = ± c2 p2z + m2 c4 . (b) Show that the corresponding eigensolutions for the wave functions are ψ2 = 0, ψ3 = a, ψ4 = 0; ψ1 = 1, ψ1 = 0, ψ2 = 1, ψ3 = 0, ψ4 = −a; ψ1 = −b, ψ2 = 0, ψ3 = 1, ψ4 = 0; ψ1 = 0, ψ2 = b, ψ3 = 0, ψ4 = 1; where a=

cpz mc2 + E+

and

b=

cpz . mc2 − E−

35. (a) Show that the simplest relativistic wave equation is pµ pµ ψ = m2 c2 ψ, where pµ = ( Ec , p) and pµ = ( Ec , −p). (b) Show also that the wave function ψ satisfies the Klein–Gordon equation   m2 c2
+ 2 ψ = 0,
where


≡ ∂µ ∂ µ ≡

 1 ∂2 2 − ∇ . c2 ∂t 2

(c) Show that the Klein–Gordon equation has plane wave solutions with positive as well as negative energy. (d) Derive ∂ ρ(r, t) = −∇ · J(r, t), ∂t where

  i
∂ψ ∂ ψ¯ ¯ ρ(r, t) = ψ −ψ 2mc2 ∂t ∂t

and J(r, t) is defined by (7.108).

This page intentionally left blank

Wavelets and Wavelet Transforms

CHAPTER

8

“Wavelets are without doubt an exciting and intuitive concept. The concept brings with it a new way of thinking, which is absolutely essential and was entirely missing in previously existing algorithms.” Yves Meyer “Multiresolution analysis provides a natural framework for the understanding of wavelet bases and for the construction of new examples. The history of the formulation of multiresolution analysis is a beautiful example of applications stimulating theoretical development.” Ingrid Daubechies

8.1 Brief Historical Remarks The concept of wavelets or ondelettes started to appear in the literature only in the early 1980s. This new concept can be viewed as a synthesis of various ideas which originated from different disciplines including mathematics (Calderón–Zygmund operators and Littlewood–Paley theory), physics (coherent states formalism in quantum mechanics and renormalization group), and engineering (quadratic mirror filters, sideband coding in signal processing, and pyramidal algorithms in image processing). In 1982, Jean Morlet, a French geophysical engineer, first introduced the idea of wavelet transform as a new mathematical tool for seismic signal analysis. It was Alex Grossmann, a French theoretical physicist, who quickly recognized the importance of the Morlet wavelet transform, which is similar to coherent states formalism in quantum mechanics, and developed an exact inversion formula for the wavelet transform. In 1984, the joint venture of Morlet and Grossmann led

433

434

Chapter 8

Wavelets and Wavelet Transforms

to a detailed mathematical study of the continuous wavelet transforms and their various applications. It has become clear from their work that, analogous to the Fourier expansions, the wavelet theory has provided a new method for decomposing a function or a signal. In many applications, especially in the time-frequency analysis of a signal, the standard Fourier transform analysis is not adequate because the Fourier transform of the signal does not contain any local information. This is probably the major weakness of the Fourier transform analysis, since it neglects the idea of frequencies changing with time or, equivalently, the notion of finding the frequency spectrum of a signal locally in time. To eliminate this weakness, in 1946 Dennis Gabor, a Hungarian–British physicist and engineer, first introduced the window-Fourier transform (or the short-time Fourier transform) or more appropriately the Gabor transform by using a Gaussian distribution function as the window function. The idea is to use a window function to localize the Fourier transform, then shift the window to another position, and so on. The remarkable difference of the Gabor transform is the local aspect of the Fourier transform analysis (with resolution in time equal to the size of the window). In fact, it deals with a discrete set of coefficients, which allows efficient numerical computation of those coefficients. However, the Gabor wavelets suffered from some serious algorithmic difficulties which were successfully resolved by Henrique Malvar in 1987. Malvar introduced new wavelets, which are now known as the Malvar wavelets and fall within the general framework of window Fourier analysis. From an algorithmic point of view, the Malvar wavelets are much more effective and superior to other wavelets including Gabor wavelets and Morlet–Grossmann wavelets. In 1985, Yves Meyer, a French pure mathematician, recognized immediately the deep connection between the Calderón formula in harmonic analysis and the new algorithm discovered by Morlet and Grossmann. Using the knowledge of the Calderón–Zygmund operators and the Littlewood–Paley theory, Meyer was able to give a mathematical foundation for wavelet theory. The first major achievement of wavelet analysis was Daubechies, Grossmann, and Meyer’s (1986) construction of a “painless” nonorthogonal wavelet expansion. During 1985 and 1986, further work of Meyer and Lemarié on the first construction of a smooth orthonormal wavelet basis on R and RN marked the beginning of their famous contributions to the wavelet theory. At the same time, Stéphans Mallat recognized that some quadratic mirror filters play an important role for the construction of orthogonal wavelet bases generalizing the Haar system. Meyer (1986) and Mallat (1988) realized that the orthogonal wavelet bases could be constructed systematically from a general formalism. Their collaboration culminated with the remarkable discovery by Mallat (1989a,b) of a new formalism, which is the so-called multiresolution analysis. It was also Mallat who constructed the wavelet decomposition and reconstruction algorithms using the multiresolution analysis. Mallat’s brilliant work was

8.1 Brief Historical Remarks

435

the major source of many new developments in wavelets. A few months later, G. Battle (1987) and Lamarié (1988) independently proposed the construction of spline orthogonal wavelets with exponential decay. While reviewing two books on wavelets in 1993, Meyer states: “Wavelets are without doubt an exciting and intuitive concept. The concept brings with it a new way of thinking, which is absolutely essential and was entirely missing in previously existing algorithms.” Mathematically, the fundamental idea of multiresolution analysis is to represent a function (or signal) f as a limit of successive approximations, each of which is a finer version of the function f . These successive approximations correspond to different levels of resolutions. Thus, multiresolution analysis is a formal approach to constructing orthogonal wavelet bases using a definite set of rules and procedures. The key feature of this analysis is to describe mathematically the process of studying signals or images at different scales. The basic principle of the MRA deals with the decomposition of the whole function space into individual subspaces Vn ⊂ Vn+1 so that the space Vn+1 consists of all rescaled functions in Vn . This essentially means a decomposition of each function (or signal) into components of different scales (or frequencies) so that an individual component of the original function f occurs in each subspace. These components can describe finer and finer versions of the original function f . For example, a function is resolved at scales t = 20 , 2−1 , . . . , 2−n . In audio signals, these scales are basically octaves, which represent higher and higher frequency components. For images and, indeed, for all signals, the simultaneous existence of a multiscale may also be referred to as multiresolution. From the point of view of practical application, MRA is really an effective mathematical framework for hierarchical decomposition of an image (or signal) into components of different scales (or frequencies). Inspired by the work of Meyer, Ingrid Daubechies (1988) made a new remarkable contribution to wavelet theory by constructing families of compactly supported orthonormal wavelets with some degree of smoothness. Her 1988 paper had a tremendous positive impact on the study of wavelets and their diverse applications. This work significantly explained the connection between the continuous wavelets on R, and the discrete wavelets on Z or ZN , where the latter has become useful for digital signal analysis. The idea of frames was introduced by Duffin and Schaeffer (1952) and subsequently studied in some detail by Daubechies (1990, 1992). In spite of tremendous success, experts in wavelet theory recognized that it is difficult to construct wavelets that are symmetric, orthogonal and compactly supported. In order to overcome this difficulty, Cohen et al. (1992a,b) studied bi-orthogonal wavelets in some detail. Chui and Wang (1991, 1992) introduced compactly supported spline wavelets and semi-orthogonal wavelet analysis. On the other hand, Beylkin, Coifman, and Rokhlin (1991) and Beylkin (1992) have successfully applied the multiresolution analysis generated by a completely orthogonal scaling function to

436

Chapter 8

Wavelets and Wavelet Transforms

study a wide variety of integral operators on L2 (R) by a matrix in a wavelet basis. This work culminated with the remarkable discovery of new algorithms in numerical analysis. Consequently, some significant progress has been made in boundary element methods, finite element methods, and numerical solutions of partial differential equations using wavelet analysis. As a natural extension of wavelet analysis, Coifman (1992a,b) in collaboration with Meyer and Wickerhauser discovered wavelet packets, which can be used to design efficient schemes for the representation and compression of acoustic signals and images. This led them to the construction of a library of orthogonal bases by extending the method of multiresolution decomposition and using the quadratic mirror filters. Recently, there have also been significant applications of wavelet analysis to a wide variety of problems in many diverse fields including mathematics, physics, medicine, computer science, and engineering. We close this historical introduction by citing some of the applications, which include addressing problems in signal processing, computer vision, seismology, turbulence, computer graphics, image processing, structures of the galaxies in the Universe, digital communication, pattern recognition, approximation theory, quantum optics, biomedical engineering, sampling theory, matrix theory, operator theory, differential equations, numerical analysis, statistics and multiscale segmentation of well logs, natural scenes, and mammalian visual systems. Wavelets allow complex information such as music, speech, images, patterns, and so on to be decomposed into elementary form, called building blocks (wavelets). The information is subsequently reconstructed with high precision. To describe the present state of wavelet research, Meyer (1993, page 31) wrote as follows: Today the boundaries between mathematics and signal and image processing have faded, and mathematics has benefited from the rediscovery of wavelets by experts from other disciplines. The detour through signal and image processing was the most direct path leading from Haar basis to Daubechies’s wavelets.

8.2 Continuous Wavelet Transforms An integral transform is an operator T on a space of functions on some  ⊂ RN which is defined by  (Tf )(y) = K(x, y)f (x)dx. 

The properties of the transform depend on the function K, which is called the kernel of the transform. For example, in the case of the Fourier transform K(x, y) = e−ixy . Note that y can be interpreted as a scaling factor. We take the exponential function ϕ(x) = e−ix and then generate a one-parameter family

8.2 Continuous Wavelet Transforms

437

of functions by taking scaled copies of ϕ, that is ϕα (x) = e−iαx , for all α ∈ R. The continuous wavelet transform is similar to the Fourier transform in the sense that it is based on a single function ψ and that this function is scaled. But unlike the Fourier transform, we also shift the function, thus generating a two-parameter family of functions ψa,b . It is convenient to define ψa,b as follows:   x−b −1/2 . ψa,b (x) = |a| ψ a Then the continuous wavelet transform is defined by  (Wψ f )(a, b) =

∞ −∞

−1/2

f (t)ψa,b (t)dt = |a|



 t−b dt. f (t)ψ a −∞ ∞



The continuous wavelet transform is not a single transform like the Fourier transform, but any transform obtained in this way. Properties of a particular transform will depend on the choice of ψ. One of the first properties we expect of any integral transform is that the original function can be reconstructed from the transform. We will prove a theorem, which gives conditions on ψ which guarantee invertibility of the transform. First we need to define the object of our study more precisely.

Definition 8.2.1. (Wavelet) By a wavelet we mean a function ψ ∈ L2 (R) satisfying the admissibility condition  ∞ 9 |ψ (ω)|2 dω < ∞, (8.1) |ω| −∞ 9 is the Fourier transform ψ , that is, where ψ  ∞ 1 9(ω) = √ e−iωx ψ(x)dx. ψ 2π −∞

If ψ ∈ L2 (R), then ψa,b (x) ∈ L2 (R) for all a, b. Indeed, 2  ∞  ∞        ψ x − b  dt = ψ(t)2 dt = ψ 2 . ψa,b (t)2 = |a|−1   a −∞ −∞ The Fourier transform of ψa,b (x) is given by    ∞ 1 x−b 9(aω). 9a,b (ω) = |a|−1/2 √ e−iωx ψ dx = |a|e−ibω ψ ψ a 2π −∞

(8.2)

(8.3)

438

Chapter 8

Wavelets and Wavelet Transforms

Figure 8.1 The Haar wavelet.

Figure 8.2 The absolute value of the Fourier transform of the Haar wavelet.

Example 8.2.2. (Haar wavelet) Let   1, ψ(x) = −1,  0,

0 ≤ x < 12 , 1 2 ≤ x < 1, otherwise

(see Figure 8.1). Then 9(ω) = √ ψ and



∞

−∞

1 (sin ω4 )2 −i(ω−π )/2 e ω (2π ) 4

9(ω)|2 8 |ψ dω = |ω| π



∞

| sin ω4 |4 dω < ∞. 3 −∞ |ω|

The Haar wavelet is one of the classic examples. It is well localized in the time domain, but it is not continuous. The absolute value of the Fourier trans9(ω)|, is plotted in Figure 8.2. This figure clearly form of the Haar wavelet, |ψ indicates that the Haar wavelet has poor frequency localization, since it does 9(ω)| is not have compact support in the frequency domain. The function |ψ even and attains its maximum at the frequency ω0 ∼ 4.662. The rate of decay as ω → ∞ is as ω−1 . The reason for the slow decay is discontinuity of ψ. Its discontinuous nature is a serious weakness in many applications. However,

8.2 Continuous Wavelet Transforms

439

the Haar wavelet is one of the most fundamental examples that illustrate major features of the general wavelet theory.


Theorem 8.2.3. Let ψ be a wavelet and let ϕ be a bounded integrable function. Then the function ψ ∗ ϕ is a wavelet.

Proof: Since 

2    ψ(x − u)ϕ(u)du dx  −∞ −∞ 2  ∞  ∞        ≤ ψ(x − u) ϕ(u) du dx

∞

−∞

 ψ ∗ ϕ(x)2 dx =



 =  ≤  ≤

∞  ∞

−∞

−∞

−∞

−∞

∞  ∞ 

−∞ ∞ −∞

 =

2     ψ(x − u)ϕ(u)1/2 ϕ(u)1/2 du dx

∞  ∞ 

−∞

  ψ(x − u)2 ϕ(u)du

 ϕ(u)du



−∞



2   ϕ(u)du

∞

−∞

∞



  ϕ(u)du dx

∞ −∞

∞ −∞

  ψ(x − u)2 ϕ(u)dx du

∞ −∞

 ψ(x)2 dx < ∞,

we have ψ ∗ ϕ ∈ L2 (R). Moreover, 

|ψ ∗ ϕ(ω)|2 dω = |ω| −∞ ∞

 

9(ω)9 |ψ ϕ (ω)|2 dω |ω| −∞ ∞

9(ω)|2  2 |ψ 9 ϕ (ω) dω |ω| −∞  9(ω)|2  2 ∞ |ψ   dω < ∞. ≤ sup 9 ϕ (ω) |ω| −∞

=

∞

Thus, the function ψ ∗ ϕ is a wavelet.




Example 8.2.4. Theorem 8.2.3 can be used to generate examples of wavelets. For example, if we take the Haar wavelet and convolve it with the following function  0, x < 0, ϕ(x) = 1, 0 ≤ x ≤ 1, 0, x ≥ 1,

440

Chapter 8

Wavelets and Wavelet Transforms

Figure 8.3 A continuous wavelet.

Figure 8.4 A smooth wavelet. then we obtain a simple continuous function, see Figure 8.3. If we convolve 2 the Haar wavelet with ϕ(x) = e−x , then the obtained wavelet is smooth, see Figure 8.4.


Definition 8.2.5.

(Continuous wavelet transform) Let ψ ∈ L2 (R) and let, for a, b ∈ R, a = 0,   x−b ψa,b (x) = |a|−1/2 ψ . a

The integral transform Wψ defined on L2 (R) by  ∞ (Wψ f )(a, b) = f (t)ψa,b (t)dt = f , ψa,b  −∞

is called a continuous wavelet transform.

(8.4)

8.2 Continuous Wavelet Transforms

441

The function ψ is often called the mother wavelet or the analyzing wavelet. The parameter b can be interpreted as the time translation and a is a scaling parameter which measures the degree of compression or scale.

Lemma 8.2.6. For any f ∈ L2 (R), we have 
9(aω). F (Wψ f )(a, b) = 2π |a| fˆ (ω)ψ

(8.5)

Proof: Using the Parseval formula for the Fourier transform, it follows from (8.4) that $ % 9a,b (Wψ f )(a, b) =  f , ψa,b  = fˆ , ψ  ∞
 1 9(aω) eibω dω. 2π |a| fˆ (ω)ψ =√ 2π −∞

(8.6)

This means that

 ∞ 
1 F (Wψ f )(a, b) = √ e−ibω (Wψ f )(a, b)db 2π −∞ 9(aω). = 2π |a| fˆ (ω)ψ

(8.7)


Theorem 8.2.7. (Parseval’s relation for wavelet transforms) Let ψ ∈ L2 (R) satisfy

 Cψ = 2π

9(ω)|2 |ψ dω < ∞. |ω| −∞ ∞

Then, for any f , g ∈ L2 (R), we have  ∞ ∞ db da (Wψ f )(a, b)(Wψ g)(a, b) 2 = Cψ f , g. a −∞ −∞

(8.8)

(8.9)

Proof: From (8.6) we get (Wψ f )(a, b) = and (Wψ g)(a, b) =





 |a| 

|a|

∞

9(aω)dω fˆ (ω)eibω ψ

(8.10)

9(aσ )dσ. gˆ (σ )e−ibσ ψ

(8.11)

−∞

∞ −∞

Substituting (8.10) and (8.11) in the left side of (8.9) gives  ∞ ∞ db da (Wψ f )(a, b)(Wψ g)(a, b) 2 a −∞ −∞

442

Chapter 8

 =

∞  −∞

Wavelets and Wavelet Transforms

∞

dbda 2 −∞ a



∞



−∞

∞ −∞

9(aω)ψ 9(aσ )eib(ω−σ ) dω dσ |a|fˆ (ω)ˆg (σ ) ψ

which is, by interchanging the order of integration,  ∞    ∞ da ∞ ∞ ˆ 1 9 9 f (ω)ˆg (σ ) ψ (aω)ψ (aσ )dω dσ = 2π eib(ω−σ ) db 2π −∞ −∞ |a| −∞ −∞  ∞   da ∞ ∞ ˆ 9(aω)ψ 9(aσ )δ(σ − ω)dω dσ f (ω)ˆg (σ ) ψ = 2π −∞ |a| −∞ −∞   ∞   da ∞ ˆ 9(aω)2 dω f (ω)ˆg (ω)ψ = 2π −∞ |a| −∞ and finally, again interchanging the order of integration and putting aω = x,  ∞  ∞ 9 2 % $ |ψ (x)| ˆ f (ω)ˆg (ω)dω dx = Cψ fˆ , gˆ = Cψ  f , g. (8.12) = 2π |x| −∞ −∞


If f = g, then (8.9) takes the form  ∞ ∞   (Wψ f )(a, b)2 db da = Cψ f 2 . a2 −∞ −∞

(8.13)

This shows that, except for the factor Cψ , the wavelet transform is an isometry from L2 (R) to L2 (R).

Theorem 8.2.8. (Inversion formula) If f ∈ L2 (R), then f (x) =

1 Cψ



∞



−∞

∞ −∞

(Wψ f )(a, b)ψa,b (x)

db da , a2

(8.14)

where the equality holds almost everywhere.

Proof: For any g ∈ L2 (R), we have 

Cψ f , g =  =

∞ −∞ ∞

 

∞ −∞

(Wψ f )(a, b)(Wψ g)(a, b) 

∞

(Wψ f )(a, b)



−∞ −∞ ∞  ∞  ∞

db da a2

∞ −∞

g(t)ψa,b (t)dt

db da a2

db da g(t)dt a2 −∞ −∞ −∞ 2 1 ∞  ∞ db da (Wψ f )(a, b)ψa,b 2 , g . = a −∞ −∞ =

(Wψ f )(a, b)ψa,b (t)

8.2 Continuous Wavelet Transforms

443

Since g is an arbitrary element of L2 (R), the inversion formula (8.14) follows.
The following theorem summarizes some elementary properties of the continuous wavelet transform. The straightforward proofs are left as exercises.

Theorem 8.2.9. Let ψ and ϕ be wavelets and let f , g ∈ L2 (R). (a) (Wψ (αf + βg))(a, b) = α(Wψ f )(a, b) + β(Wψ g)(a, b) for any α, β ∈ C, (b) (Wψ (Tc f ))(a, b) = (Wψ f )(a, b − c), where Tc is the translation operator defined by Tc f (t) = f (t − c), (c) (Wψ (Dc f ))(a, b) =

√1 (Wψ f )( a , b ), c c c

where c is a positive number and Dc

is the dilation operator defined by Dc f (t) = 1c f ( ct ), (d) (Wψ ϕ)(a, b) = (Wϕ ψ)( 1a , − ab ), a = 0, (e) (Wαψ+βϕ f )(a, b) = α(Wψ f )(a, b) + β(Wϕ f )(a, b) for any α, β ∈ C, (f) (WPψ Pf )(a, b) = (Wψ f )(a, −b), where P is the parity operator defined by Pf (t) = f (−t), (g) (WTc ψ f )(a, b) = (Wψ f )(a, b + ca), (h) (WDc ψ f )(a, b) =

√1 (Wψ f )(ac, b), c

c > 0.

For the wavelets to be useful analyzing functions, the mother wavelet must have certain properties. One such property is defined by the condition (8.1) which guarantees existence of the inversion formula for the con9 is tinuous wavelet transform. If ψ ∈ L1 (R), then its Fourier transform ψ 9 is continuous, Cψ can be finite only if ψ 9(0) = 0, or equivcontinuous. If ψ ∞ alently, −∞ ψ(t)dt = 0. This means that ψ must be an oscillatory function with zero mean. Condition (8.1) also imposes a restriction on the rate of de9(ω)|2 . cay of |ψ In addition to the admissibility condition (8.1), there are other properties that may be useful in particular applications. For example, we may want to require that ψ be n times continuously differentiable or infinitely differentiable. If the Haar wavelet is convolved n + 1 times with the function ϕ given in Example 8.2.4, then the resulting function ψ ∗ ϕ ∗ · · · ∗ ϕ is an n times differentiable wavelet. The function in Figure 8.4 is an infinitely differentiable wavelet. The so-called “Mexican hat wavelet” is another example of an infinitely differentiable wavelet.

Example 8.2.10. (Mexican hat wavelet) This wavelet is defined by   2 ψ(t) = 1 − t 2 e−t /2 and shown in Figure 8.5.




444

Chapter 8

Wavelets and Wavelet Transforms

Figure 8.5 Mexican hat wavelet. Another desirable property of wavelets is the so-called “localization property.” We want ψ to be well localized in both time and frequency. In other words, ψ and its derivatives must decay very rapidly. For frequency localiza9(ω) must decay sufficiently rapidly as ω → ∞ and ψ 9(ω) should be flat tion, ψ in the neighborhood of ω = 0. The flatness at ω = 0 is associated with the number of vanishing moments of ψ. The kth moment of ψ is defined by  ∞ mk = t k ψ(t)dt. −∞

A wavelet is said to have n vanishing moments if  ∞ t k ψ(t)dt = 0 for k = 0, 1, . . . , n, −∞

or, equivalently, ,

9(ω) dk ψ dωk

= 0 for k = 0, 1, . . . , n. ω=0

Wavelets with a larger number of vanishing moments result in more flatness when frequency ω is small.

8.3 The Discrete Wavelet Transform While the continuous wavelet transformis compared to the Fourier transform, ∞ which requires calculating the integral −∞ e−iωx f (x)dx for all (or almost all) ω ∈ R, the discrete wavelet transform can  2π be compared to the Fourier series, which requires calculating the integral 0 e−inx f (x)dx for integer values of n.

8.3 The Discrete Wavelet Transform

445

Since the continuous wavelet transform is a two-parameter representation of a function    ∞ t−b −1/2 dt, f (t)ψ (Wψ f )(a, b) = |a| a −∞ we could discretize it by assuming that a and b take only integer values. It turns out that it is better to discretize it in a different way. First we fix two positive constants a0 and b0 and then define  −m/2  ψm,n (x) = a0 ψ a−m (8.15) 0 x − nb0 , where m and n range over Z. By the discrete wavelet coefficients of f ∈ L2 (R) we mean the numbers f , ψm,n , where m, n ∈ Z. The fundamental question here is whether it is possible to reconstruct f from those coefficients. The weakest interpretation of this problem is whether f , ψm,n  = g, ψm,n  for all m, n ∈ Z implies f = g. In practice we expect much more than that: we want f , ψm,n  and g, ψm,n  to be “close” if f and g are “close.” This will be guaranteed if there exists a B > 0, such that ∞    f , ψm,n 2 ≤ B f 2 m,n=−∞

for all f ∈ L2 (R). Similarly, we want f and g to be “close” if f , ψm,n  and g, ψm,n  are “close.” This is important because we want to be sure that when we neglect some small terms in the representation of f in terms of f , ψm,n , then the reconstructed function will not differ much from f . The representation will have this property if there exists an A > 0, such that A f 2 ≤

∞    f , ψm,n 2 m,n=−∞

for all f ∈ L2 (R). These two requirements are best studied in terms of the socalled frames.

Definition 8.3.1. (Frame) A sequence (ϕ1 , ϕ2 , . . .) in a Hilbert space H is called a frame if there exist A, B > 0 such that ∞    2 f , ϕn 2 ≤ B f 2 A f ≤ (8.16) n=1

for all f ∈ H . The constants A and B are called frame bounds. If A = B, then the frame is called tight.

446

Chapter 8

Wavelets and Wavelet Transforms

 2 If (ϕn ) is an orthonormal basis, then it is a tight frame since√ ∞ n=1 |f , ϕn | √ 3 3 1 1 2 = f for all f ∈ H. The vectors (1, 0), (− 2 , 2 ), (− 2 , − 2 ) form a tight frame in C2 which is not a basis. As pointed out above, we want the family of functions ψm,n to form a frame in L2 (R). (Obviously, the double indexing of the functions is irrelevant.) The following theorem gives fairly general, sufficient conditions for (ψm,n ) to constitute a frame in L2 (R).

Theorem 8.3.2. If ψ and a0 are such that ∞    m 2 ψ 9 a0 s  > 0,

inf

1≤|s|≤a0

m=−∞ ∞    m 2 ψ 9 a s  0 and some constant C, then there exists b˜ such that (ψm,n ) is a ˜ frame in L2 (R) for any b0 ∈ (0, b).

Proof: Let f ∈ L2 (R). Then  ∞    

∞    f , ψm,n 2 = m,n=−∞

m,n=−∞

 ∞    

=

m,n=−∞

∞ −∞ ∞ −∞

2  −m/2 m f (x)a0 ψ(a0 x − nb0 )dx 2  m ib a ns m 0 ˆf (s)am/2 ψ 0 9(a0 s)e ds = ♣ 0

by the general Parseval relation (Theorem 5.11.21), basic properties of the Fourier transform (Theorem 5.11.7), and  ∞the fact that we sum over all integers. Since, for any ω > 0, the integral −∞ g(t)dt can be written as ∞  

ω

g(t + lω)dt,

l=−∞ 0

by taking ω = ♣=

2π , b 0 am 0

∞  m,n=−∞

we obtain

2  ∞    ω     9 am am e2π ins/ω fˆ (s + lω)ψ 0 (s + lω) ds 0   0 l=−∞

8.3 The Discrete Wavelet Transform

=

 

∞ 

 am 0 

m,n=−∞



ω

e

2π ins/ω

0

∞  l=−∞

447

 2    m ˆf (s + lω)ψ 9 a0 (s + lω) ds 

2  ω   ∞ ∞       9 am = am fˆ (s + lω)ψ  0 (s + lω)  ds = ♦, 0ω   0 m=−∞

l=−∞

by Parseval’s formula for trigonometric Fourier series. Since 2  ∞    m  ˆ 9 a0 (s + lω)  f (s + lω)ψ    l=−∞  ∞  ∞     m  m   ˆ ˆ 9 a0 (s + kω) 9 a0 (s + lω) f (s + lω)ψ f (s + kω)ψ = l=−∞

k=−∞

and ∞ 

F(s) =

  9 am fˆ (s + kω)ψ 0 (s + kω)

k=−∞

is a periodic function with a period of ω, we have    m 9 a0 (s + lω) F(s)ds fˆ (s + lω)ψ

 ω  ∞ 0

l=−∞

 =

∞ −∞

  9 am fˆ (s)ψ 0 s F(s)ds

∞  

=

∞

k=−∞ −∞

    ˆ 9 m 9 am fˆ (s)ψ 0 s f (s + kω)ψ a0 (s + kω) ds.

Consequently, 2π ♦= b0 2π = b0 +

 ∞ 

∞

m,k=−∞ −∞



∞   m 2   ˆf (s)2 ψ 9 a0 s  ds

∞

−∞

2π b0

    m 9 9 am fˆ (s)fˆ (s + kω) ψ 0 s ψ a0 (s + kω) ds

m=−∞

 m,k=−∞ k=0



∞ −∞

    m 9 9 am fˆ (s)fˆ (s + kω) ψ 0 s ψ a0 (s + kω) ds.

448

Chapter 8

Wavelets and Wavelet Transforms

To find a bound on the second summand, we use the Schwarz inequality:   2π  b 0





−∞

m,k=−∞ k=0

≤

2π b0

∞

    m   ˆf (s)fˆ (s + kω) ψ  9 am 9 0 s ψ a0 (s + kω) ds

 

m,k=−∞ k=0

 ×

    m   m  1/2 ˆf (s)2 ψ 9 a0 s ψ 9 a0 (s + kω) ds

∞ −∞

    m   m  1/2 ˆf (s + kω)2 ψ 9 a0 (s + kω) ds 9 a0 s ψ = ♥.

∞ −∞

Then, by first changing the variables in the second factor and by using Hölder’s inequality (Theorem 1.2.7), we have 2π ♥= b0

  m,k=−∞ k=0

 × ≤

    m   m  1/2 ˆf (s)2 ψ 9 a0 (s + kω) ds 9 a0 s ψ

∞ −∞

   m   m  1/2 ˆf (s)2 ψ 9 a0 s ds 9 a0 (s − kω) ψ

∞ −∞

 ∞  ∞ ∞    m   m    1/2 2π  ˆf (s)2 ψ 9 a0 s ψ 9 a0 (s + kω) ds b0 −∞ m=−∞ k=−∞ k=0

 ×

 ∞   m     m  1/2 ˆf (s)2 ψ 9 a0 s ds 9 a0 (s − kω) ψ = ♠.

∞ −∞

m=−∞

If we denote β(ξ ) = sup

∞    m   m  ψ 9 a s ψ 9 a s + ξ ,

s∈R m=−∞

0

0

then ∞    m   m 1/2 2π 2 f β a0 kω β −a0 kω ♠≤ b0 k=−∞ k=0

  ∞ ,   2π 2π k 2π k 1/2 2 = f . β β − b0 b0 b0 k=−∞ k=0

449

8.3 The Discrete Wavelet Transform

Consequently, if we denote    -1/2  ∞ ∞ ,      m 2 2π k 2π k 2π ψ 9 a0 s  − sup β β − A= b0 s∈R m=−∞ b0 b0 k=−∞ k=0

and

   -1/2  ∞ ∞ ,      m 2 2π k 2π k 2π ψ 9 a0 s  + inf , β β − B= b0 s∈R m=−∞ b0 b0 k=−∞ k=0

we obtain A f 2 ≤

∞    f , ψm,n 2 ≤ B f 2 . m,n=−∞

−(1+ε)

, we have Since β(ξ ) ≤ C(1 + |ξ |)     ∞ ,  ∞ ,    2π k 2π k 1/2 2π k 2π k 1/2 =2 β β − β β − b0 b0 b0 b0 k=−∞ k=0

k=1

≤ 2C

∞   k=1

∞

 ≤ 2C

0

=

2π k 1+ b0 1+

−(1+ε)

2π t b0

−(1+ε) dt

Cb0 . πε

∞ 0 9 m 2 Now, since Cb m=−∞ |ψ (a0 s)| > 0, π ε → 0 as b0 → 0 and inf1≤|s|≤a0 ˜ Fithere exists a positive  number b˜ such that A > 0 for any b0 ∈ (0, b). m 2 9 nally, since sup1≤|s|≤a0 ∞ | ψ (a s)| < ∞, we also have B < ∞ for all 0 m=−∞ ˜ Thus, ψm,n constitute a frame for all such b0 . b0 ∈ (0, b).
Now we turn to the problem of reconstruction of f from f , ψm,n  and representation of f in terms of ψm,n . For a complete  orthonormal system (ϕn ), both questions are answered by the equality f = ∞ n=1 f , ϕn ϕn . As we will see, since we do not have orthogonality, the situation is more complicated for frames.

Definition 8.3.3.

(Frame operator) Let (ϕ1 , ϕ2 , . . .) be a frame in a Hilbert space H . The operator F from H into l2

450

Chapter 8

Wavelets and Wavelet Transforms

defined by

  F(f ) = f , ϕn 

is called a frame operator.

Lemma 8.3.4. Let F be a frame operator. Then F is a bounded, linear, and invertible operator. Its inverse F −1 is also a bounded operator.

The proof is easy and is left as an exercise. Consider the adjoint operator F ∗ of the frame operator F associated with a frame (ϕn ). For any (cn ) ∈ l2 , we have (∞ ) ∞  % $ %  $ ∗ cn ϕn , f  = cn ϕn , f . F (cn ), f = (cn ), Ff = n=1

n=1

Thus, the adjoint operator of a frame operator has the form F ∗ (cn ) =

∞ 

cn ϕn .

(8.17)

n=1

Note also that since ∞    $ %  f , ϕn 2 = Ff 2 = F ∗ Ff , f , n=1

the condition (8.16) can be expressed as A I ≤ F ∗ F ≤ BI , where the inequality ≤ is to be understood in the sense defined in Section 4.6.

Theorem 8.3.5. Let (ϕ1 , ϕ2 , . . .) be a frame with frame bounds A and B and let F be the associated frame operator. Define  −1 ϕ˜ n = F ∗ F ϕn . Then (ϕ˜ n ) is a frame with frame bounds

1 B

and A1 .

Proof: By Corollary 4.5.6, we have (F ∗ F)−1 = ((F ∗ F)−1 )∗ . Consequently, $  −1 % $ −1 %  f , ϕ˜ n  = f , F ∗ F ϕn = F ∗ F f , ϕn and then ∞ ∞   $ $ ∗ −1 % %  f , (ϕ˜ n ) 2 =  F F f , ϕ n 2 n=1

n=1

8.3 The Discrete Wavelet Transform

451

  −1 2 = F F ∗ F f  $  −1  −1 % = F F∗F f , F F∗F f $ −1 % = F∗F f , f . Now, since AI ≤ F ∗ F ≤ BI , Theorem 4.6.11 implies  −1 1 1 I ≤ F∗F ≤ I, B A which gives ∞  $ % 1  f , (ϕ˜ n ) 2 ≤ 1 f 2 . f 2 ≤ B A n=1




The sequence (ϕ˜ n ) is called the dual frame.

Lemma 8.3.6. Let F be the frame operator associated with the frame (ϕ1 , ϕ2 , . . .), and let F˜ be the frame operator associated with the dual frame (ϕ˜1 , ϕ˜ 2 , . . .). Then ˜ F˜ ∗ F = I = F ∗ F.

Proof: Since  −1 $ −1  %  ˜ , F F ∗ F f = F ∗ F f , ϕn =  f , ϕ˜ n  = Ff we have

  −1 ∗  −1 F = F∗F F∗F = I F˜ ∗ F = F F ∗ F

and  −1 F ∗ F˜ = F ∗ F F ∗ F = I.




Now we are ready to state and prove the theorem which answers the question of reconstructability of f from (f , ϕn ).

Theorem 8.3.7. Let (ϕ1 , ϕ2 , . . .) be a frame in a Hilbert space H and let (ϕ˜ 1 , ϕ˜ 2 , . . .) be the dual frame. Then ∞  f=  f , ϕn ϕ˜ n n=1

and f=

∞   f , ϕ˜n ϕn n=1

for any f ∈ H.

452

Chapter 8

Wavelets and Wavelet Transforms

Proof: Let F be the frame operator associated with (ϕn ) and let F˜ be the

frame operator associated with the dual frame (ϕ˜n ). Since I = F˜ ∗ F, for any f ∈ H, we have ∞    ∗ ∗ ˜ ˜ f = F Ff = F  f , ϕn  =  f , ϕn ϕ˜n , n=1

by (8.17). The proof of the other equality is similar.




8.4 Multiresolution Analysis and Orthonormal Bases of Wavelets Some difficulties in dealing with frames arise from lack of orthogonality. If we have orthogonality, that is, ψk, , ψm,n  = 0 whenever (k, ) = (m, ∞n), then the reconstruction of f from f , ψm,n  is much simpler: f = m,n=−∞ f , ψm,n ψm,n . In this section we describe a general method of constructing orthonormal bases of wavelets based on the so-called multiresolution analysis. In this construction we take a0 = 2 and b0 = 1. Consequently,   ϕm,n (x) = 2−m/2 ϕ 2−m x − n .

Definition 8.4.1.

(Multiresolution analysis) By a multiresolution analysis we mean a sequence (. . . , V−2 , V−1 , V0 , V1 , V2 , . . .) of spaces of functions defined on R such that the following conditions are satisfied: (a) Vn is a closed subspace of L2 (R) for every n ∈ Z, (b) Vn+1 ⊂ Vn for every n ∈ Z,

2 (c) ∞ n=−∞ Vn is dense in L (R), # (d) ∞ n=−∞ Vn = {0}, (e) f ∈ Vn if and only if f (2n ·) ∈ V0 , for all n ∈ Z, (f) There exists a ϕ ∈ V0 such that {ϕ0,k : k ∈ Z} is an orthonormal basis in V0 . The function ϕ is called the scaling function or father wavelet. The fundamental idea of multiresolution analysis is to represent a function as a limit of successive approximations, each of which is a “smoother” version of the original function. These approximations correspond to different resolutions. Multiresolution analysis is a formal approach to constructing orthogonal wavelet bases using a definite set of rules and procedures. In applications,

8.4 Multiresolution Analysis

453

it is an effective mathematical framework for hierarchical decomposition of a signal or an image into components of different scales. The spaces Vn correspond to those different scales.

Example 8.4.2. Let ϕ be the characteristic function of the interval [0, 1]. Define

 Vn =

∞ 

 ck ϕn,k : (ck ) ∈ l (Z) . 2

k=−∞

Then (Vn ) is a multiresolution analysis. This example is discussed in more detail in the next section.
The spaces in a multiresolution analysis form a ladder: · · · ⊃ V−2 ⊃ V−1 ⊃ V0 ⊃ V1 ⊃ V2 ⊃ · · · . If any of the spaces Vn is given, then, by condition (e), any other space Vm is known, because
   Vm = f 2m−n · : f ∈ Vn . In most arguments concerning multiresolution analysis, it thus is enough to deal with V0 . Notice that a resolution analysis is completely determined by the scaling function ϕ. Indeed, for a given ϕ, we first define  ∞   V0 = ck ϕ0,k : (ck ) ∈ l2 (Z) . k=−∞

The remaining spaces Vn are then determined by V0 as discussed earlier. This does not mean that any choice of ϕ is going to produce a multiresolution analysis. The first problem we have to deal with is orthogonality of translates of ϕ. The following theorem can be useful in that matter.

Theorem 8.4.3. For any ϕ ∈ L2 (R), the following two conditions are equivalent: (a) The system {ϕ0,k : k ∈ Z} is orthonormal, (b)

∞

ϕ (ω + 2mπ )|2 m=−∞ |9

= 1/2π almost everywhere.

Proof: Since ϕ (ω), 9 ϕ0,k (ω) = e−ikω 9 we have ϕ0,0 , 9 ϕ0,l−k  ϕ0,k , ϕ0,l  = ϕ0,0 , ϕ0,l−k  = 9  ∞  2 ϕ (ω) dω e−i(l−k)ω 9 = −∞

454

Chapter 8

Wavelets and Wavelet Transforms

=

∞  

2(m+1)π

 2 ϕ (ω) dω e−i(l−k)ω 9

m=−∞ 2mπ



2π

=

e−i(l−k)ω

0

∞   2 9 ϕ (ω + 2mπ ) dω. m=−∞

Thus, ϕ0,k , ϕ0,l  = δk,l if and only if most everywhere.

∞

ϕ (ω m=−∞ |9

+ 2mπ )|2 = 1/2π al


Now we return to the problem of constructing an orthonormal basis of wavelets from a given multiresolution analysis. Let (Vn ) be a multiresolution analysis. For every n ∈ Z, define Wn to be the orthogonal complement of Vn in Vn−1 , so that we have Vn−1 = Vn ⊕ Wn

Wn ⊥ Wm

and

if m = n.

Since, Vn = Vn+1 ⊕ Wn+1 = Vn+2 ⊕ Wn+2 ⊕ Wn+1 = Vn+3 ⊕ Wn+3 ⊕ Wn+2 ⊕ Wn+1 = ···, for every N > n, we have Vn = VN ⊕

N−n−1 >

 WN−k .

k=0

This, in view of (c) and (d) of Definition 8.4.1, leads to L (R) = 2

∞ >

Wn .

(8.18)

n=−∞

We have thus constructed a decomposition of L2 (R) into mutually orthogonal subspaces. Moreover, for all n ∈ Z,   f ∈ Wn if and only if f 2n · ∈ W0 . This is a very desirable property because if {θk : k ∈ Z} is an orthonormal basis for W0 , then the scaled version {θk (2n ·): k ∈ Z} is an orthonormal basis for Wn , for every n ∈ Z. This fact, together with (8.18), implies that the union of all these bases, that is {θk (2n ·): k, n ∈ Z}, is an orthonormal basis for L2 (R). Note that the spaces Vn do not have this property. This is the reason for defining the spaces Wn .

8.4 Multiresolution Analysis

455

The property just described reduces the problem of finding an orthonormal basis for L2 (R), to the problem of finding an orthonormal basis for W0 . We want to go one step further: we want to find a single function ψ such that the translates {ψ0,n : n ∈ Z} constitute an orthonormal basis in W0 . If we can find a function like that, then the system {ψm,n : m, n ∈ Z}, will be an orthonormal basis for L2 (R). It will be an orthonormal basis of wavelets. It turns out that such a ψ exists for every multiresolution analysis. The proof of this fact not only guarantees existence of such a function, but actually gives a method of constructing it. The function ψ is often called the mother wavelet. We start the construction by introducing an auxiliary function m 9, which plays an important role in the argument. Since ϕ ∈ V−1 and (ϕ−1,n ) is an orthonormal basis for V−1 , we can express ϕ in the form ϕ=

∞ 

hn ϕ−1,n

or

ϕ(x) =

n=−∞

√

2

∞ 

hn ϕ(2x − n)

(8.19)

n=−∞

where hn = ϕ, ϕ−1,n 

and

∞ 

|hn |2 = 1.

(8.20)

n=−∞

By applying the Fourier transform to (8.19), we obtain ∞ 1  hn e−ins/2 9 ϕ (s/2). 9 ϕ (s) = √ 2 n=−∞

(8.21)

∞ 1  hn e−ins , m 9(s) = √ 2 n=−∞

(8.22)

If we define

then (8.21) can be written as 9 ϕ (s) = m 9(s/2)9 ϕ (s/2).

(8.23)

∞   2 1 9 ϕ (s + 2kπ ) = 2π

(8.24)

By Theorem 8.4.3, we have

k=−∞

which, together with (8.23), gives ∞   2  2 1 m 9(s + kπ ) 9 ϕ (s + kπ ) = 2π

k=−∞

a.e.

(8.25)

456

Chapter 8

Wavelets and Wavelet Transforms

By splitting the sum into even and odd k and then using 2π -periodicity of m 9, we obtain ∞   2  2 m 9(s + kπ ) 9 ϕ (s + kπ )

(8.26)

k=−∞

=

∞   2  2 m 9(s + 2kπ ) 9 ϕ (s + 2kπ )

(8.27)

k=−∞

+

∞    2   2 m 9 s + (2k + 1)π  9 ϕ s + (2k + 1)π 

(8.28)

k=−∞

=

∞ ∞    2  2  2  2 m m 9(s) 9 ϕ (s + 2kπ ) + 9(s + π ) 9 ϕ (s + π + 2kπ ) (8.29) k=−∞

k=−∞

2  2  1  9(s) + m , = m 9(s + π ) 2π

(8.30)

by (8.24) used in its original form and with s replaced by s + π . Now (8.25) implies  2  2 m 9(s) + m 9(s + π ) = 1 a.e. (8.31) The Fourier transform 9 ϕ of the scaling function ϕ satisfies the functional equation (8.23). The function m 9 is called the generating function of the multiresolution analysis. This function is often called the discrete Fourier transform of the sequence (hn ). In signal processing, m 9(ω) is called the transfer function of a discrete filter with impulse response (hn ) or the low-pass filter associated with the scaling function ϕ. Since |9 ϕ (0)| = 1 = 0, we have m 9(0) = 1 and m 9(π ) = 0. Thus m 9 can be considered as a low-pass filter because the transfer function passes the frequencies near s = 0 and cuts off frequencies near ω = π .

Theorem 8.4.4. The function 9 ϕ can be represented by the infinite product   ∞ s m 9 k . 9 ϕ (s) = 2 k=1

Proof: A simple iteration of (8.23) gives  ,        s s s s s 9 ϕ =m 9 m 9 9 ϕ 9 ϕ (s) = m 9 2 2 2 4 4

(8.32)

8.4 Multiresolution Analysis

which gives, by the (k − 1)th iteration,         s s s s m 9 ...m 9 k 9 ϕ k 9 ϕ (s) = m 9 2 4 2 2     k s s m 9 k 9 ϕ k . = 2 2

457

(8.33)

k=1

Since 9 ϕ (0) = 1 and 9 ϕ (s) is continuous, we obtain   s ϕ k =9 ϕ (0) = 1. lim 9 k→∞ 2 Thus, the limit of (8.33) as k → ∞ gives (8.32).




Now we prove a technical lemma.

Lemma 8.4.5. For every f ∈ W0 there exists a 2π -periodic function ν such that fˆ (s) = ν(s)eis/2 m 9(s/2 + π ) 9 ϕ (s/2).

(8.34)

9(s/2 + π ) 9 ϕ (s/2) in (8.34) is independent of f . Note that eis/2 m

Proof: Let f ∈ W0 . Since W0 ⊂ V−1 , we can write f=

∞ 

cn ϕ−1,n ,

n=−∞

where cn = f , ϕ−1,n . Proceeding as before, we get ∞  ˆf (s) = √1 cn e−ins/2 9 ϕ (s/2) = m 9f (s/2)9 ϕ (s/2), 2 n=−∞

(8.35)

∞ 1  m 9f (s) = √ cn e−ins . 2 n=−∞

(8.36)

where

Since f ⊥ V0 ,



∞ −∞

and hence



2π

e 0

ins

∞  k=−∞

ϕ (s)eins ds = 0, fˆ (s)9

fˆ (s + 2kπ )9 ϕ (s + 2kπ )ds = 0,

458

Chapter 8

Wavelets and Wavelet Transforms

for every n ∈ Z. Consequently, ∞ 

fˆ (s + 2kπ )9 ϕ (s + 2kπ ) = 0.

(8.37)

k=−∞

Now we substitute (8.23) and (8.35) in (8.37), then split the sum into even and odd k’s, and finally use 2π -periodicity of m 9 and m 9f : ∞ 

0=

ϕ (s + 2kπ ) fˆ (s + 2kπ )9

k=−∞ ∞ 

=

 2 m 9f (s/2 + kπ )9 m(s/2 + kπ )9 ϕ (s/2 + kπ )

k=−∞ ∞ 

=

 2 m 9f (s/2 + 2kπ )9 m(s/2 + 2kπ )9 ϕ (s/2 + 2kπ )

k=−∞

+

∞ 

 2 m 9f (s/2 + π + 2kπ )9 m(s/2 + π + 2kπ )9 ϕ (s/2 + π + 2kπ )

k=−∞ ∞ 

=

 2 m 9f (s/2)9 m(s/2)9 ϕ (s/2 + 2kπ )

k=−∞

+

∞ 

 2 m 9f (s/2 + π )9 m(s/2 + π )9 ϕ (s/2 + π + 2kπ )

k=−∞

=m 9f (s/2)9 m(s/2)

∞   2 9 ϕ (s/2 + 2kπ ) k=−∞

+m 9f (s/2 + π )9 m(s/2 + π )

∞   2 9 ϕ (s/2 + π + 2kπ ) k=−∞

  1 . = m 9f (s/2)9 m(s/2) + m 9f (s/2 + π )9 m(s/2 + π ) 2π The last equality follows from the fact that the system {ϕ0,k : k ∈ Z} is orthonormal and from Theorem 8.4.3. The obtained result and (8.37) imply m 9f (s)9 m(s) + m 9f (s + π )9 m(s + π ) = 0 a.e. Thus,

   m m 9(s + π )   9f (s) = 0,  −9 9(s)  mf (s + π ) m

8.4 Multiresolution Analysis

459

which can be interpreted as linear dependence of the vectors (9 mf (s), m(s + π ), m 9(s)). Consequently, there exists a function λ −9 mf (s + π )), and (9 such that m 9f (s) = λ(s)9 m(s + π )

a.e.

(8.38)

Since both m 9 and m 9f are 2π -periodic, so is λ. Moreover, λ(s) + λ(s + π ) = 0 a.e.

(8.39)

There exists a 2π -periodic function ν defined by λ(s) = eis ν(2s).

(8.40)

Combining (8.35), (8.38), and (8.40), we obtain the desired representation (8.34).
Now we return to our main problem. Suppose that we found a function ψ such that {ψ0,n : n ∈ Z} is a basis for W0 . Then every function f ∈ W0 has a representation ∞ 

f= where obtain

∞

γn ψ0,n =

n=−∞

2 n=−∞ |γn |

∞ 

γn ψ(x − n),

(8.41)

n=−∞

< ∞. When we apply the Fourier transform to (8.41), we 

fˆ (s) = ∞

∞ 

 9(s) = γˆ (s)ψ 9(s). γn e−ins ψ

n=−∞

The function γˆ (s) = n=−∞ γn e−ins is a 2π -periodic and square integrable on [0, 2π ]. When this is compared with (8.34), then we see that we should take     s s 9(s) = eis/2 m ψ 9(s/2 + π )9 ϕ (s/2) = m 91 9 ϕ , (8.42) 2 2 where the function m 91 is given by 9(s + π ). m 91 (s) = eis m

(8.43)

This function m 91 is called the filter conjugate to m 9 and hence, in signal processing, m 9 and m 91 are called conjugate quadratic filters (CQF). Since 9(s) = eis/2 m 9(s/2 + π )9 ϕ (s/2) ψ ∞ 1  hn ein(s/2+π ) 9 ϕ (s/2) = eis/2 √ 2 n=−∞

460

Chapter 8

Wavelets and Wavelet Transforms ∞ 1  =√ hn einπ ei(n+1)(s/2) 9 ϕ (s/2), 2 n=−∞

by letting n = −(k + 1), we obtain     ∞ s iks 1  k−1 9(s) = √ 9 ϕ . h−k−1 (−1) exp − ψ 2 2 2 n=−∞

(8.44)

Invoking the inverse Fourier transform to (8.44), with k replaced by n, gives the mother wavelet ψ(x) =

∞ √  2 (−1)n−1 h−n−1 ϕ(2x − n),

(8.45)

n=−∞

=

∞ √  2 dn ϕ(2x − n),

(8.46)

n=−∞

where the coefficients dn are given by dn = (−1)n−1 h−n−1 .

(8.47)

Thus, the representation (8.46) of a mother wavelet has the same structure as that of the father wavelet ϕ given by (8.19). The mother wavelet ψ associated with a given multiresolution analysis is not unique because the coefficients dn = (−1)n−1 h2N−1−n

(8.48)

define the same mother wavelet (8.45) with suitably selected n ∈ Z. The wavelet with coefficients dn given by (8.48) has the Fourier transform      is s s 9 ψ (s) = exp (2N − 1) m 9 +π 9 ϕ . (8.49) 2 2 2 Thus, the nonuniqueness property of dn allows us to define a form of ψ different from (8.45) by ψ(x) =

√

∞ 

2

dn ϕ(2x − n),

(8.50)

n=−∞

with a slightly modified dn given by dn = (−1)n h1−n .

(8.51)

In practice, any one of the preceding formulas for dn can be used to find a mother wavelet.

461

8.4 Multiresolution Analysis

The orthogonality condition (8.31), together with (8.23) and (8.42), implies     2  2  s 2  9    9(s) = 9 ϕ (s) + ψ (8.52) ϕ 2  . Or, equivalently,

  m 2   m 2  2 9 9 2 s  = 9 ϕ 2 s  + ψ ϕ (s) .

(8.53)

Summing both sides of (8.53) from m = 1 to infinity leads to the result ∞   m 2  2  ψ 9  92 s . ϕ (s) =

(8.54)

m=1

If ϕ has a compact support, the series (8.46) for the mother wavelet ψ terminates and, consequently, ψ can be represented by a finite linear combination of a translated version of ϕ(2x).

Theorem 8.4.6. Let (Vn ), n ∈ Z be a multiresolution analysis with the scaling function ϕ. Then the function ψ(x) =

∞ √  2 (−1)n−1 h−n−1 ϕ(2x − n), n=−∞

where hn =

√  2

∞ −∞

ϕ(x)9 ϕ (2x − n)dx,

is a mother wavelet, that is, the system {ψm,n : m, n ∈ Z} is an orthonormal basis for L2 (R).

Proof: First we need to verify that {ψ0,n : m, n ∈ Z} is an orthonormal set. Indeed, we have  ∞  ∞   9(s)2 ds 9 ψ(x − k)ψ (x − l)dx = e−is(k−l) ψ −∞



−∞ 2π

= 0

e−is(k−l)

∞    ψ 9(s + 2nπ )2 ds, n=−∞

and ∞ ∞       2 2  ψ m 9(s + 2nπ )2 = 9 s/2 + (n + 1)π  9 ϕ (s/2 + nπ ) n=−∞

n=−∞

=

∞    2 2  m 9 s/2 + (2n + 1)π  9 ϕ (s/2 + 2nπ ) n=−∞

462

Chapter 8

Wavelets and Wavelet Transforms ∞    2    m 9 s/2 + (2n + 1)π 2 9 s/2 + (2n + 2)π  φ + n=−∞ ∞  2  2  9 ϕ (s/2 + 2nπ ) = m 9(s/2 + π ) n=−∞ ∞    2  2 9 ϕ s/2 + (2n + 1)π  + m 9(s/2) n=−∞

2  2  1  1 9(s/2 + π ) + m = . = m 9(s/2) 2π 2π Thus,



∞ −∞

9(x − l)dx = δk,l , ψ(x − k)ψ

which proves that {ψm,n : m, n ∈ Z} is an orthonormal system. In view of Lemma 8.4.5 and the discussion following the lemma, to prove that it is a basis it suffices to show that the function ν in (8.40) is square integrable over [0, 2π ]. In fact, 

2π 

 ν(s)2 ds = 2

0



π

 λ(s)2 ds

0



2    2  λ(s)2 m 9(s + π ) + m 9(s) ds

π

=2 0



2π 

 2 λ(s)2 m 9(s + π ) ds

=2 0



2π 

2 m 9f (s) ds

=2 0

= 2π

∞ 

|cn |2

(where cn = f , ϕ−1,n )

n=−∞

= 2π f 2 < ∞.




8.5 Examples of Orthonormal Wavelets Since the discovery of wavelets, orthonormal wavelets play an important role in the wavelet theory and have a variety of applications. In this section, we discuss several examples of orthonormal wavelets.

8.5 Examples of Orthonormal Wavelets

463

Definition 8.5.1.

(Orthonormal wavelet) A wavelet ψ ∈ L (R) is called orthonormal if the family of functions ψm,n generated from ψ by      n ψm,n (x) = 2m/2 ψ 2m x − m = 2m/2 ψ 2m x − n , m, n ∈ Z, (8.55) 2 2

is orthonormal, that is,  ψm,n , ψk,  =

∞ −∞

ψm,n (x)ψk, (x)dx = δm,k δn, ,

(8.56)

for all m, n, k,  ∈ Z. The following lemma is often useful when dealing with orthogonality of wavelets.

Lemma 8.5.2. If ψ, ϕ ∈ L2 (R), then ψm,k , ϕm,  = ψn,k , ϕn, ,

(8.57)

for all m, n, k,  ∈ Z.

Proof: We have  ψm,k , ϕm,  =

∞ −∞

    2m ψ 2m x − k ϕ 2m x −  dx,

which is, by letting 2m x = 2n t,  ∞     ψm,k , ϕm,  = 2n ψ 2n t − k ϕ 2n t −  dx = ψn,k , ϕn, . −∞




Example 8.5.3. (The Haar wavelet) The simplest example of an orthonormal wavelet is the classic Haar wavelet. Consider the scaling function ϕ = χ[0,1) . The function ϕ satisfies the dilation equation ϕ(x) =

∞ √  2 cn ϕ(2x − n),

(8.58)

n=−∞

where the coefficients cn are given by √  ∞ ϕ(x)ϕ(2x − n)dx. cn = 2 −∞

(8.59)

464

Chapter 8

Wavelets and Wavelet Transforms

Figure 8.6 Two-scale relation of ϕ(x) = ϕ(2x) + ϕ(2x − 1). Evaluating this integral with ϕ = χ[0,1) gives cn as follows: 1 c0 = c1 = √ 2

and

cn = 0

for n > 1.

Consequently, the dilation equation becomes ϕ(x) = ϕ(2x) + ϕ(2x − 1).

(8.60)

This means that ϕ(x) is a linear combination of the even and odd translates of ϕ(2x) and satisfies a very simple two-scale relation (8.60), as shown in Figure 8.6. In view of (8.51), we obtain 1 d0 = c1 = √ 2

and

1 d1 = −c0 = − √ . 2

Thus, the Haar mother wavelet is obtained from (8.50) as a simple two-scale relation ψ(x) = ϕ(2x) − ϕ(2x − 1) = χ[0,1/2] (x) − χ[1/2,1] (x)  1 if 0 ≤ x < 12 , = −1 if 1 ≤ x < 1, 2  0 otherwise.

(8.61)

(8.62)

This two-scale relation (8.61) of ψ is represented in Figure 8.7. For any m, n ∈ Z, we have   2−m/2 ,  ψm,n (t) = 2−m/2 ψ 2−m t − n = −2−m/2 ,  0, 

2m n ≤ t < 2m n + 2m−1 , 2m n + 2m−1 ≤ t < 2m n + 2m , otherwise.

8.5 Examples of Orthonormal Wavelets

465

Figure 8.7 Two-scale relation of ψ(x) = ϕ(2x) − ϕ(2x − 1). Clearly, ψm,n 2 = ψ 2 = 1, for all m, n ∈ Z. To verify that {ψm,n } is an orthonormal system, note that  ψm,n , ψk,  =

∞ −∞

    2m/2 ψ 2m x − n 2k/2 ψ 2k x −  dx,

which gives, by the change of variables 2m x − n = t, 

∞

  ψ(t)ψ 2k−m (t + n) −  dt.

(8.63)

ψ(t)ψ(t + n − )dt = δ0,n− = δn, ,

(8.64)

ψm,n , ψk,  = 2k/2 2−m/2

−∞

For m = k, we obtain  ψm,n , ψm,  =

∞ −∞

where ψ(t) = 0 in 0 ≤ t < 1 and ψ(t −  − n) = 0 in  − n ≤ t < 1 +  − n, and these intervals are disjoint from each other unless n = . We now consider the case m = k. In view of symmetry, it suffices to consider the case m > k. Putting r = m − k > 0 in (8.63), we obtain, for m > k,  ψm,n , ψm,  = 2

r/2

∞ −∞

  ψ(t)ψ 2r t + s dt,

where s = 2r n − . Thus, it suffices to show that 

1/2 0



 ψ 2 t + s dt − r



1 1/2

  ψ 2r t + s dt = 0.

(8.65)

466

Chapter 8

Wavelets and Wavelet Transforms

Invoking a simple change of variables 2r t + s = x, we find 

1/2

  ψ 2r t + s dt −



0



 ψ 2 t + s dt =

1



r

1/2

a

 ψ(x)dx −

s

b

ψ(x)dx, (8.66) a

where a = s + 2r−1 and b = s + 2r . Since the interval [s, a] contains the support [0, 1] of ψ, the first integral in (8.66) is zero. Similarly, the second integral is also zero.


Example 8.5.4. (The Shannon wavelet) The function ψ whose Fourier transform satisfies 9(ω) = χI (ω), ψ

(8.67)

where I = [−2π, −π ] ∪ [π, 2π ], is called the Shannon wavelet. The function 9 so that ψ can be directly obtained from the inverse Fourier transform of ψ ψ(t) = =

1 2π 1 2π



∞ −∞

,

9(ω)dω eiωt ψ

−π −2π



2π

eiωt dω +

eiωt dω

π

  sin( π2t ) 1 3π t (sin 2π t − sin π t) = π t cos . = πt 2 (2)

(8.68)

This function is orthonormal to its translates by integers. Indeed, by Parseval’s relation, $ % % 1 $ 9 9, einω ψ ψ(t), ψ(t − n) = ψ 2π  ∞ 1 9(ω)einω ψ 9(ω)dω = ψ 2π −∞  2π 1 einω dω = δ0,n . = 2π −2π The wavelet basis is now given by   1 , ψm,n (t) = 2−m/2 ψ 2−m t − n − 2 or sin{ π (2−m t − n)} ψm,n (t) = 2−m/2 π 2 −m t − n) 2 (2

 cos

m, n ∈ Z

 3π  −m 2 t−n . 2

(8.69)

8.5 Examples of Orthonormal Wavelets

467

Figure 8.8 The Shannon wavelet and its Fourier transform. For any fixed n ∈ Z, the functions ψm,n (t) form a basis for the space of functions supported on the interval  −m+1    −2 π, −2−m π ∪ 2−m π, 2−m+1 π . The system {ψm,n (t)}, m, n ∈ Z, is an orthonormal basis for L2 (R). Both ψ(t) 9(ω) are shown in Figure 8.8. and ψ The Fourier transform of ψm,n is  m/2 m −m −m+1 π, 9m,n (ω) = 2 exp(−iωn2 ) if 2 π < |ω| < 2 (8.70) ψ 0 otherwise. 9m,n and ψ 9k, do not overlap for m = k. Hence, by Theorem 5.11.21, Evidently, ψ it turns out that, for m = k, ψm,n , ψk,  =

% 1 $ 9k, = 0. 9m,n , ψ ψ 2π

(8.71)

For m = k, we have % 1 $ 9m, 9m,n , ψ ψ 2π 
  −m 2 1 −m ∞ 9 2 ω  dω = exp −iω2−m (n − ) ψ 2 2π −∞  ∞
 1 exp −iσ (n − ) dσ = δn, . (8.72) = 2π −∞

ψm,n , ψk,  =

This shows that {ψm,n (t)} is an orthonormal system.




Example 8.5.5. (The Daubechies wavelets and algorithms) Daubechies (1988, 1992) first developed the theory and construction of continuous orthonormal

468

Chapter 8

Wavelets and Wavelet Transforms

wavelets with compact support. Wavelets with compact support have many interesting properties. They can be constructed to have a given number of derivatives and to have a given number of vanishing moments. We assume that the scaling function ϕ satisfies the dilation equation ϕ(x) =

√

∞ 

2

cn ϕ(2x − n),

(8.73)

n=−∞

 2 where cn = ϕ, ϕ1,n  and ∞ n=−∞ |cn | ≤ ∞. If the scaling function ϕ has compact support, then only a finite number of cn have nonzero values. The associated generating function m 9, ∞ 1  m 9(ω) = √ cn e−iωn 2 n=−∞

(8.74)

is a trigonometric polynomial and it satisfies the identity (8.31) with special values m 9(0) = 1 and m 9(π ) = 0. If coefficients cn are real, then the corresponding scaling function, as well as the mother wavelet ψ, will also be real-valued. The mother wavelet ψ corresponding to ϕ is given by the formula (8.42) with 9(ω) is m-times continuously differen|9 ϕ (0)| = 1. The Fourier transform ψ tiable and it satisfies the moment condition 9(k) (0) = 0 for k = 0, 1, . . . , m. ψ

(8.75)

It follows that ψ ∈ C m (R) implies that m 90 has a zero at ω = π of order (m + 1). In other words,   1 + e−iω m+1 9 m 90 (ω) = L(ω), (8.76) 2 where 9 L is a trigonometric polynomial. In addition to the orthogonality condition (8.31), we assume 

1 + e−iω m 90 (ω) = 2

N 9 L(ω),

(8.77)

where 9 L(ω) is 2π -periodic and 9 L ∈ C N−1 (R). Evidently, 2  m 90 (ω)9 m0 (−ω) 90 (ω) = m     1 + e−iω N 1 + eiω N 9 L(ω)9 L(−ω) = 2 2 N   2 2 ω 9 L(ω) , = cos 2

(8.78)

(8.79)

8.5 Examples of Orthonormal Wavelets

469

where |9 L(ω)|2 is a polynomial in cos ω, that is,  2 9 L(ω) = Q(cos ω). Since cos ω = 1 − 2 sin2 ( ω2 ), it is convenient to introduce x = sin2 ( ω2 ) so that (8.79) reduces to the form  N  2 2 ω m  90 (ω) = cos Q(1 − 2x) = (1 − x)N P(x), (8.80) 2 where P(x) is a polynomial in x. We next use the fact that     ω+π ω cos2 = sin2 =x 2 2 and

 2 9 L(ω + π ) = Q(− cos ω) = Q(2x − 1),   = Q 1 − 2(1 − x) = P(1 − x)

(8.81)

to express the identity (8.31) in terms of x so that (8.31) becomes (1 − x)N P(x) + xN P(1 − x) = 1.

(8.82)

Since (1 − x)N and xN are two polynomials of degree N, which are relatively prime, then, by Bezout’s theorem (see Daubechies, 1992), there exists a unique polynomial PN of degree ≤ N − 1 such that (8.82) holds. An explicit solution for PN (x) is given by PN (x) =

N−1  k=0

 N +k−1 k x, k

(8.83)

which is positive for 0 < x < 1 so that PN (x) is at least a possible candidate for |9 L(ω)|2 . There also exist higher degree polynomial solutions PN (x) of (8.82) which can be written as   N−1 N + k − 1 1 k N PN (x) = x +x R x− , (8.84) k 2 k=0

where R is an odd polynomial. L(ω)|2 and Since PN (x) is a possible candidate for |9  2 9 L(ω)9 L(−ω) = 9 L(ω) = Q(cos ω) = Q(1 − 2x) = PN (x),

(8.85)

the next problem is how to find 9 L(ω). This can be done by the following lemma.

470

Chapter 8

Wavelets and Wavelet Transforms

Lemma 8.5.6. (Riesz’s lemma for spectral factorization) If 9 A(ω) =

n 

ak cosk ω,

(8.86)

k=0

where ak ∈ R and an = 0, and if 9 A(ω) ≥ 0 for all ω ∈ R with 9 A(0) = 1, then there exists a trigonometric polynomial 9 L(ω) =

n 

bk e−ikω

(8.87)

k=0

L(0) = 1 and with real coefficients bk such that 9  2 9 A(ω) = 9 L(ω)9 L(−ω) = 9 L(ω)

(8.88)

for all ω ∈ R. We refer to Daubechies (1992) for a proof of the Riesz lemma 8.5.6. We also point out that the factorization of 9 A(ω) given in (8.88) is not unique. For a given N, if we select P = PN , then 9 A(ω) becomes a polynomial of 9 degree N − 1 in cos ω and L(ω) is a polynomial of degree (N − 1) in exp(−iω). Therefore, the generating function m 90 given by (8.77) is of degree (2N − 1) in exp(−iω). The interval [0, 2N − 1] becomes the support of the corresponding scaling function N ϕ. The mother wavelet N ψ obtained from N ϕ is called the Daubechies wavelet. For N = 2, it follows from (8.83) that P2 (x) =

 1   k+1 k=0

k

xk = 1 + 2x

and hence (8.85) gives    2 2 ω 2 ω 9  L (ω) = P2 (x) = P2 sin = 1 + 2 sin2 = 2 − cos ω. 2 2 Using (8.87) in Lemma 8.5.6, we obtain that 9 L(ω) is a polynomial of degree N − 1 = 1 and  1 9 L(ω)9 L(−ω) = 2 − eiω + e−iω . 2 It follows from (8.87) that     1 b0 + b1 e−iω b0 + b1 eiω = 2 − eiω + e−iω . 2

(8.89)

8.5 Examples of Orthonormal Wavelets

471

Equating the coefficients in this identity gives b20 + b21 = 1

and

2b0 b1 = −1.

(8.90)

These equations admit solutions b0 =

√  1 and 1+ 3 2

b1 =

√  1 1− 3 . 2

(8.91)

Thus, the generating function (8.77) takes the form    1 + e−iω 2  m 90 (ω) = b0 + b1 e−iω 2 ; < √   √  √  √    1  1 + 3 + 3 + 3 e−iω + 3 − 3 e−2iω + 1 − 3 e−3iω = 8 (8.92) with m 90 (0) = 1. Comparing coefficients of (8.92) with (8.22) gives hn = cn as √  1  c0 = √ 1 + 3 , 4 2 √  1  c2 = √ 3 − 3 , 4 2

√  1  c1 = √ 3 + 3 , 4 2 √  1  c3 = √ 1 − 3 . 4 2

(8.93)

Consequently, the Daubechies scaling function 2 ϕ(x) takes the form, dropping the subscript, √   ϕ(x) = 2 c0 ϕ(2x) + c1 ϕ(2x − 1) + c2 ϕ(2x − 2) + c3 ϕ(2x − 3) . (8.94) Using (8.48) with N = 2, we obtain the Daubechies wavelet 2 ψ(x), dropping the subscript, √   ψ(x) = 2 d0 ϕ(2x) + d1 ϕ(2x − 1) + d2 ϕ(2x − 2) + d3 ϕ(2x − 3) √   = 2 −c3 ϕ(2x) + c2 ϕ(2x − 1) − c1 ϕ(2x − 2) + c0 ϕ(2x − 3) , (8.95) where the coefficients in (8.95) are the same as for the scaling function ϕ(x), but in reverse order and with alternate terms having their signs changed from plus to minus. On the other hand, the use of (8.46) with (8.51) also gives the Daubechies wavelet 2 ψ(x) in the form √   2 ψ(x) = 2 −c0 ϕ(2x − 1) + c1 ϕ(2x) − c2 ϕ(2x + 1) + c3 ϕ(2x + 2) . The wavelet has the same coefficients as ψ given in (8.95) except that the wavelet is reversed in sign and runs from x = −1 to 2 instead of starting from

472

Chapter 8

Wavelets and Wavelet Transforms

Figure 8.9 The Daubechies scaling function 2 ϕ(x).

Figure 8.10 The Daubechies wavelet 2 ψ(x). x = 0. It is often referred to as the Daubechies D4 wavelet since it is generated by four coefficients. Both Daubechies’ scaling function 2 ϕ and Daubechies’ wavelet 2 ψ are shown in Figures 8.9 and 8.10 respectively.
For N = 1, it follows from (8.83) that P1 (x) ≡ 1, and this in turn leads to the fact that Q(cos ω) = 1, 9 L(ω) = 1 so that the generating function is m 90 (ω) =

 1 1 + e−iω . 2

This corresponds to the generating function for the Haar wavelet.

(8.96)

8.6 Exercises

473

8.6 Exercises 1. If f is a homogeneous function of degree n, show that (Wψ f )(λa, λb) = λn+1/2 (Wψ f )(a, b). 2. In the proof of Theorem 8.2.7 we do not address the difficulty that arises from the fact that a takes both positive and negative values. Provide a more detailed proof that removes the difficulty. 3. Prove Theorem 8.2.9. 4. Prove that the vectors (1, 0), (− 12 , in C.

√ √ 3 3 1 2 ), (− 2 , − 2 )

form a tight frame

5. Prove that if (ϕn ) is a tight frame in a Hilbert space H with frame bound A, then ∞  Af , g =  f , ϕn ϕn , g n=1

for all f , g ∈ H. 6. Prove that if (ϕn ) is a tight frame in a Hilbert space H with frame bound 1, then (ϕn ) is an orthonormal basis in H. 7. Show that 

∞

1 πn sin π x sin π(2x − n) dx = sin . π(2x − n) 2π n 2 −∞ π x

8. Let ψm,n (t) = 2−m/2 ψ(2−m t − n), where ψ(t) is the Haar wavelet, that is,  −m/2 if 2m n < t < 2m n + 2m−1 , 2 ψm,n (t) = −2−m/2 if 2m n + 2m−1 < t < 2m n + 2m , 0 otherwise and   a if 0 < t < 12 , f (t) = b if 1 < t < 1, 2  0 otherwise. (a) Find f , ψm,0 . (b) Show that

∞

m=0 f , ψm,0 ψm,0 (t) =



a if 0 < t < 12 , b if 12 < t < 1.

474

Chapter 8

Wavelets and Wavelet Transforms

9. For the Shannon wavelet ψ(t) =

sin( π2t )

ψm,n (t) = 2

πt 2 − m2

 cos

 3π t , 2

   sin( π2 (2−m t − n)) 3π  −m cos 2 t−n , π −m t − n) 2 2 (2

show that ψm,n (t) are orthonormal wavelets. 10. The cardinal B-splines Bn (x) of order n are defined by the following convolution product Bn (x) = B1 (x) ∗ B1 (x) ∗ · · · ∗ B1 (x) = B1 (x) ∗ Bn−1 (x) (n ≥ 2), where n factors are involved in the convolution product and B1 (x) = χ[0,1] (x). Show the following: x (a) Bn (x) = x−1 Bn−1 (t)dt. (b) B2 (x) = xχ[0,1] (x) + (2 − x)χ[1,2] (x). (c) B3 (x) =

x2 2 χ[0,1]

+ 12 (6x − 2x2 − 3)χ[1,2] + 12 (x − 3)2 χ[2,3] .

(d) B4 (x) = 16 x3 χ[0,1] + 13 (2 − 6x + 6x2 − x3 )χ[1,2] + 12 (x3 − 2x2 + 20x − 13)χ[2,3] . 11. Use Exercise 10 to show that the two-scale relation for the B-splines of order n is   n  1−n n Bn (2x − k). Bn (x) = 2 k k=0

12. Use the Fourier transform 9 B1 (ω) of B1 (x) to prove ∞   2 sin2n (ω) d2n−1 9 Bn (2ω + 2π k) = − (cot ω). (2n − 1)! dω2n−1

k=−∞

13. The Franklin wavelet is generated by the second-order (n = 2) splines. Show that the Fourier transform 9 ϕ (ω) of this wavelet is   sin2n ω2 2 2 ω −1/2 9 ϕ (ω) = ω 2 1 − sin . (2) 3 2 Hence, calculate the Fourier transform of the Franklin wavelet ϕ.

8.6 Exercises

475

14. The harmonic wavelet ψ(t) due to Newland (1993, 1994) is defined by 9(ω) = ψ 9e (ω) + iψ 90 (ω), where its Fourier transform ψ  1 9 ψe (ω) = 4π if −4π ≤ ω < −2π and 2π ≤ ω < 4π , 0 otherwise, and

  4πi 9 ψ0 (ω) = − i  4π 0

if −4π ≤ ω < −2π , if 2π ≤ ω < 4π , otherwise.

Show that 9(t) = ψe (t) + iψ0 (t) = ψ

 1  exp(4π it) − exp(2π it) , 2π it

where ψe and ψ0 are real even and odd functions respectively and the 9(ω) is defined by using the factor 1 . Fourier transform ψ 2π 15. (a) Show that the family of harmonic wavelets ψ(2m t − k) forms an orthogonal set. (b) Show that



∞ −∞

  ψ 2m t − k 2 dt = 2−m .

This page intentionally left blank

Optimization Problems and Other Miscellaneous Applications

CHAPTER

9

“Physics can’t exist without mathematics which provides it with the only language in which it can speak. Thus, services are continuously exchanged between pure mathematical analysis and physics. It is really remarkable that among works of analysis most useful for physics were those cultivated for their own beauty. In exchange, physics, exposing new problems, is as useful for mathematics as it is a model for an artist.” Henri Poincaré “In every mathematical investigation, the question will arise whether we can apply our mathematical results to the real world.” V.I. Arnold

9.1 Introduction The fundamental problems in classical functional analysis include the following: (a) The existence and uniqueness of solutions for operator equations in a Hilbert space (Tx = y, where T is a bounded linear operator from a Hilbert space H1 into a Hilbert space H2 and y is a given element of H2 ). (b) The minimization of the norm in a Hilbert space, the so-called minimum norm optimization problems. (c) The maximization or minimization of a functional on a Hilbert space, the so-called variational problems.

477

478

Chapter 9

Miscellaneous Applications

In previous chapters, the first problem has been discussed in some detail as an application of the theory of operators on Hilbert spaces. This chapter is essentially concerned with the optimization and variational problems and their applications. To discuss these problems, it is necessary to introduce some ideas of the calculus of operators in Banach spaces. The rest of the chapter deals with applications to miscellaneous problems, including minimization of a quadratic functional, variational inequalities, and optimal control problems for dynamical systems. Also included are approximation theory, the Shannon sampling theorem, linear and nonlinear stability problems, and bifurcation theory.

9.2 The Gateaux and Fréchet Differentials Throughout this section, B1 and B2 are Banach spaces over a field F, which may be either R, the real numbers, or C, the complex numbers. Assume that T : B1 → B2 is an operator (not necessarily linear) with the domain D(T) = B1 .

Definition 9.2.1.

(Gateaux differential) Suppose x is a fixed element of B1 . The operator T : B1 → B2 is said to be Gateaux differentiable at x if there exists a continuous linear operator A such that     T(x + th) − T(x) − A(h) lim (9.1) =0  t→0 t

for every h ∈ B1 , where t → 0 in F. The operator A is called the Gateaux differential of T at x, and its value at h is denoted by A(h) = dT(x, h).

According to the definition, the Gateaux differential of an operator from B1 into B2 at x ∈ B1 is a linear operator from B1 into B2 . Note that if T is a linear operator, then dT(x, h) = T(h), that is, dT(x) = T for all x ∈ B1 . If f is a functional on B1 , that is, f : B1 → F, and f is Gateaux differentiable at some x ∈ B1 , then , d f (x + th) df (x, h) = dt t=0 and, for each fixed x ∈ B1 , df (x, h) is a linear functional of h ∈ B1 .

Theorem 9.2.2. If the Gateaux differential exists, it is unique.

9.2 The Gateaux and Fréchet Differentials

479

Proof: Suppose that two operators A1 and A2 satisfy (9.1). Then, for every h ∈ B1 and every t > 0, we have   A1 (h) − A2 (h)       T(x + th) − T(x) T(x + th) − T(x)  (h) − (h) = − A − A 2 1   t t      T(x + th) − T(x)   T(x + th) − T(x)  ≤ + − A2 (h) − A1 (h)     → 0, t t

as t → 0. Therefore, A1 (h) − A2 (h) = 0 for all h ∈ B1 , proving the theorem.


Definition 9.2.3. (Gradient of a functional) The mapping x → df (x, ·) is called the gradient of f , and is usually denoted by ∇f . Consequently, ∇ is a mapping from B1 into the dual space B 1 . Example 9.2.4. Let B1 = RN and let f be a functional on B1 , that is, f : B1 →

R. Let x = (x1 , . . . , xN ) ∈ B1 and h = (h1 , . . . , hN ) ∈ B1 . If f has continuous partial derivatives of order one, then the Gateaux differential of f is df (x, h) =

N  ∂f (x) k=1

∂xk

hk .

For a fixed x0 ∈ B1 , the Gateaux differential at x0 , ' & N  ∂f (x) df (x0 , h) = hk ∂xk k=1

(9.2)

,

(9.3)

x=x0

is a bounded linear operator from RN into RN . (Note that, in this example, B 1 = B1 .) We can also write 1 df (x0 , h) =

 2 ∂f (x0 ) ∂f (x0 ) ,..., ,h . ∂x1 ∂xN




Example 9.2.5. Let B1 = RN and B2 = RM . Let f = (f1 , . . . , fM ) : B1 → B2 be Gateaux differentiable at some x ∈ RN . The Gateaux differential A can be identified with an M × N matrix (aij ). If h is the jth coordinate vector, h = ej =

480

Chapter 9

Miscellaneous Applications

(0, . . . , 1, . . . , 0), then

    f (x + th) − f (x) − A(h) lim =0  t→0 t

implies

   fi (x + tej ) − fi (x)   lim − aij  = 0, t→0 t

for every i = 1, . . . , M and j = 1, . . . , N. This shows that fi ’s have partial derivatives at x and ∂fi (x) = aij , ∂xj for every i = 1, . . . , M and j = 1, . . . , N. The Gateaux differential of f at x has the matrix representation  ∂f (x) ∂f1 (x)  1 ...  ∂x1 ∂xN   .. ..    (9.4) . .  = (aij ).   ∂f (x)  ∂fM (x) M ... ∂x1 ∂xN This is called the Jacobian matrix of f at x. Note that if M = 1 then the matrix reduces to a row vector, which is the case discussed in Example 9.2.4.


Example 9.2.6. Let B = C ([a, b]) be the normed space of real-valued continuous functions on [a, b] with the norm defined by   x = sup x(t). t∈[a,b]

Let K(s, t) be a continuous real valued function defined on [a, b] × [a, b], and let g(t, x) be a continuous real valued function on [a, b] × R with continuous partial derivative ∂g/∂x on [a, b] × R. Define a mapping f : B → B by  b   f (x)(s) = K(s, t)g t, x(t) dt. (9.5) a

Then

, df (x, h) =

d dα



b a

  K(s, t)g t, x(t) + αh(t) dt

. α=0

Interchange of the order of differentiation and integration is permissible under the given assumption on g, and hence it follows that ,  b  ∂  g t, x(t) h(t)dt. (9.6) K(s, t) df (x, h) = ∂x a

9.2 The Gateaux and Fréchet Differentials

481

Thus, the Gateaux differential of the integral operator (9.5) is the linear integral operator (9.6) and its kernel is K(s, t)gx (t, x).
Note that the Gateaux differential is a generalization of the idea of the directional derivative familiar in finite dimensional spaces.

Theorem 9.2.7. (Mean value theorem) Suppose the functional f has a Gateaux differential df (x, h) at every point x ∈ B. Then, for any two points x, x + h ∈ B, there exists a ξ ∈ (0, 1) such that f (x + h) − f (x) = df (x + ξ h, h).

(9.7)

Proof: Put (t) = f (x + th). Then ,

, (t + s) − (t) f (x + th + sh) − f (x + th)  (t) = lim = lim s→0 s→0 s s

= df (x + th, h). Application of the mean value theorem for functions of one variable to  yields (1) − (0) =  (ξ ) for some ξ ∈ (0, 1). Consequently, f (x + h) − f (x) = df (x + ξ h, h).




The derivative of a function f of a real variable is defined by f (x + h) − f (x) , h→0 h

f (x) = lim

(9.8)

provided the limit exists. This definition cannot be used in the case of mappings defined on a Banach space because h is then a vector, and division by a vector is meaningless. On the other hand, division by a vector can be easily avoided by rewriting (9.8) as f (x + h) = f (x) + f (x)h + hw(h),

(9.9)

where w is a function (which depends on h) such that w(h) → 0 as h → 0. Equivalently, we can now say that f (x) is the derivative of f at x if f (x + h) − f (x) = f (x)h + (h),

(9.10)

where (h) = hw(h), and thus (h)/h → 0 as h → 0, usually denoted by (h) = o(h). The definition based on (9.10) can be generalized to include mappings from a Banach space into a Banach space. This leads to the concept of the Fréchet differentiability and Fréchet derivative (Maurice René Fréchet (1878–1973)).

482

Chapter 9

Miscellaneous Applications

Definition 9.2.8.

(Fréchet differential) Let x be a fixed point in a Banach space B1 . A continuous linear operator A : B1 → B2 is called the Fréchet differential of the operator T : B1 → B2 at x if T(x + h) − T(x) = Ah + (x, h)

(9.11)

and lim

h →0

(x, h) =0 h

(9.12)

or, equivalently, T(x + h) − T(x) − Ah = 0. h →0 h lim

(9.13)

The Fréchet differential at x will be denoted by T (x) or dT(x). In the case of a real-valued function f : R → R, the ordinary derivative at x is a number representing the slope of the graph of the function at x. The Fréchet differential of f is not a number, but a linear operator from R into R. The existence of the ordinary derivative f (x) implies the existence of the Fréchet differential at x, and the comparison of (9.9) and (9.11) shows that A is the operator, which multiplies every h ∈ R by the number f (x). In elementary calculus, the tangent to a curve is the straight line giving the best approximation of the curve in the neighborhood of the point of tangency. Similarly, the Fréchet differential of an operator f can be interpreted as the best local linear approximation. We consider the change in f when its argument changes from x to x + h, and then we approximate this change by a linear operator A so that f (x + h) = f (x) + Ah + ε,

(9.14)

where ε is the error in the linear approximation. Thus, ε has the same order of magnitude as h, except for the case when A is equal to the Fréchet derivative of f . In such a case, ε = o(h), so that ε is much smaller than h as h → 0. In this sense, the Fréchet differential gives the best linear approximation of f near x. Finally, if A is a linear operator, then the derivative of A is A itself and the best linear approximation of A is A itself.

Theorem 9.2.9. If a mapping has the Fréchet differential at a point, then it has the Gateaux differential at that point and both differentials are equal. Proof: Let T : B1 → B2 , and let x ∈ B1 . If T has the Fréchet differential at x, then T(x + h) − T(x) − Ah =0 h →0 h lim

9.2 The Gateaux and Fréchet Differentials

483

for some continuous linear operator A : B1 → B2 . In particular, for any fixed nonzero h ∈ B1 , we have       T(x + th) − T(x)  T(x + th) − T(x) − A(th)    h = 0.   lim − Ah = lim  t→0 t→0 t th Thus, A is the Gateaux differential of T at x.




Corollary 9.2.10. If the Fréchet differential exists, it is unique. Proof: Suppose A1 and A2 are Fréchet differentials of T at some x ∈ B1 . Then A1 and A2 are the Gateaux differentials of T at x. Thus, A1 = A2 , by Theorem 9.2.2.
Example 9.2.11. Let f : R2 → R be defined by  

x3 y f (x, y) = x4 + y2  0

if x = 0 and y = 0, if x = y = 0.

It is easy to check that f is Gateaux differentiable at 0, and the Gateaux differential at that point is 0. On the other hand, since |x3 x2 | 1 1 1 |f (x, x2 )| = = √ → √ 2 4 4 2 4 2 (x, x ) (x + x ) x + x 2 2 1+x f is not Fréchet differentiable at (0,0).

as x → 0,


Example 9.2.12. Suppose f : H → F is a functional on a Hilbert space H, where F = R or C. If f is Fréchet differentiable at some x ∈ H, then that differential must be a bounded linear functional on H, that is, f (x) ∈ H . By the Riesz Representation Theorem 3.7.7, there exists an element y ∈ H such that f (x)(h) = h, y for every h ∈ H. The differential f (x) can thus be identified with the vector y, which is called the gradient of f , ∇f .
Example 9.2.13. Let T be a bounded linear operator on a real Hilbert space H, and let f be a functional on H defined by f (x) = x, Tx. We will show that the Fréchet differential of f is f (x)(h) = h, (T + T ∗ )x. Let h be an arbitrary nonzero element of H. Then $   % f (x + h) − f (x) − h, T + T ∗ x $ % $ % = x + h, T(x + h) − x, Tx − h, Tx − h, T ∗ x $ % = x, Tx + x, Th + h, Tx + h, Th − x, Tx − h, Tx − h, T ∗ x = h, Th.

484

Chapter 9

Miscellaneous Applications

Consequently, f (x + h) − f (x) − h, T + T ∗ x |h, Th| = lim = 0. h →0 h →0 h h lim

So the result follows.




Example 9.2.14. By the Hammerstein operator T : C ([a, b]) → C ([a, b]), we mean an operator defined by 

b

(Tu)(x) =

  K(x, t)f t, u(t) dt,

a

where K : [a, b] × [a, b] → R and f : [a, b] × R → R are given functions. If f is sufficiently smooth, then 

b

T(u + h)(x) = a

, 1 2 K(x, t) f (t, u) + hfu (t, u) + h fuu (t, u) + · · · dt 2

= (Tu)(x) + Ah + o(h), where the Fréchet differential A = T (u) is



b

T (u)(h) =

  K(x, t)fu t, u(t) h(t)dt.

a

Thus, the Fréchet derivative of T at u is the linear integral operator with the kernel K(x, t)fu (t, u(t)).


Theorem 9.2.15. If an operator defined on an open subset of a Banach space is Fréchet differentiable at a point, then it is continuous at that point. Proof: Let  be an open set in a Banach space B1 , and let T be an operator from  into a Banach space B2 . Let x ∈  and let ε > 0 be such that x + h ∈  whenever h < ε. Then     T(x + h) − T(x) = Ah + (x, h) → 0 as h → 0. This proves that T is continuous at x.




Much of the theory, results, and methods of ordinary calculus can be easily generalized to Fréchet derivatives. For example, the usual rules for differentiation of the sum and product (in the case of functionals) of two or more functions apply to Fréchet differentials. The mean value theorem, the implicit function theorem, and Taylor series have satisfactory extensions. The interested reader is referred to Liusternik and Sobolev (1974). In the next theorem, we prove the chain rule for Fréchet differentials.

9.2 The Gateaux and Fréchet Differentials

485

Theorem 9.2.16. (Chain rule) Let B1 , B2 , and B3 be real Banach spaces. If g : B1 → B2 is Fréchet differentiable at some x ∈ B1 and f : B2 → B3 is Fréchet differentiable at y = g(x) ∈ B2 , then  = f ◦ g is Fréchet differentiable at x and    (x) = f g(x) g (x).

Proof: For x, h ∈ B1 , we have     (x + h) − (x) = f g(x + h) − f g(x)   = f g(x + h) − g(x) + g(x) − f (y) = f (d + y) − f (y), where d = g(x + h) − g(x). Thus,     (x + h) − (x) − f (y)d = o d . In view of d − g (x)h = o( h ), we obtain       (x + h) − (x) − f (y)g (x)h = o h + o d . Since g is continuous at x, by Theorem 9.2.15, we have d = o( h ) and thus    (x)h = f g(x) g (x)h.


Theorem 9.2.17. A linear operator T from a Banach space into a Banach space is Fréchet differentiable if and only if T is bounded. In that case, T = T. Proof: If T is a linear operator and Fréchet differentiable at a point, then T is continuous (and hence bounded), by Theorem 9.2.15. Conversely, if T is a bounded linear operator, then   T(x + h) − Tx − Th = 0, proving that T is Fréchet differentiable and T = T.




Definition 9.2.18.

(Second Fréchet differential) If T : B1 → B2 is Fréchet differentiable on an open set  ⊂ B1 and T is Fréchet differentiable at x ∈ , then T is called twice Fréchet differentiable at x. The Fréchet differential of T at x is called the second Fréchet differential of T and is denoted by T

(x). Note that if T : B1 → B2 is Fréchet differentiable on an open set  ⊂ B1 , then T is a mapping from B1 into B (B1 , B2 ) (recall that B (B1 , B2 ) denotes the

486

Chapter 9

Miscellaneous Applications

space of all bounded linear mappings from B1 into B2 ). Consequently, if T

(x) exists, it is a bounded linear mapping from B1 into B (B1 , B2 ). If T

exists at every point of , then T

: B1 → B (B1 , B (B1 , B2 )). In ordinary calculus, the analogue of the formula (9.9) for the second derivative can be written as   1 f (x + h) = f (x) + f (x)h + f

(x)h2 + o h2 , 2

(9.15)

where o(h2 ) stands for a function ϕ(h) = w(h)h2 such that limh→0 w(h) = 0. Clearly, (9.15) is the truncated Taylor expansion of f (x + h) in powers of h. It says that f (x) is the coefficient of the linear term h and 12 f

(x) is the coefficient of the quadratic term h2 . Thus, (9.15) can be used to find f

in the same way that (9.9) is used to calculate f . To extend (9.15) to the case of the second Fréchet differential of a scalar valued function on a Banach space, it seems reasonable to interpret f

(x)h2 as the result of application of the linear operator f

(x) to the pair (h, h), and o(h2 ) as a function ϕ such that ϕ(h) / h 2 → 0 as h → 0.

Example 9.2.19. To calculate f

(x) for a differentiable real-valued function f : RN → R, we use Taylor series expansion f (x + h) = f (x) +

 ∂f 1  ∂2 (x)hi + (x)hi hj + · · · . ∂xi 2 ∂xi ∂xj i

(9.16)

i,j

Thus, f (x) represents the gradient operator defined by f (x)h =

 ∂f (x)hi . ∂xi i

The second Fréchet differential f

(x) is defined by f

(x)(h, h) =

 i,j

 ∂2 (x)hi hj = fij (x)hi hj , ∂xi ∂xj

(9.17)

i,j

where fij denotes the second partial derivative. In other words, f

(x, x) is the bilinear operator given by the matrix of the second partial derivative of f . For a twice differentiable function f : RN → R, we have fij = fji , so the matrix is symmetric. This leads to the following result for a general pair (h, k):  f

(x)(h, k) = fij (x)hi kj . (9.18) i,j

In general, we can define the second Fréchet differential f

(x) as a symmetric bilinear operator. For a symmetric bilinear mapping ϕ, ϕ(h, k) is uniquely

9.2 The Gateaux and Fréchet Differentials

487

determined by its values for ϕ(h, h). In fact, we have ϕ(h + k, h + k) = ϕ(h, h) + 2ϕ(h, k) + ϕ(k, k),

(9.19)

and hence, ϕ(h, k) =

 1 ϕ(h + k, h + k) − ϕ(h, h) − ϕ(k, k) . 2

(9.20)

This formula can be used to obtain an expression for f

(x)(h, k) in terms of f

(x)(h, h).


Example 9.2.20. Consider a nonlinear integral operator T : C ([a, b]) → C ([a, b]) defined by



b

(Tu)(t) =

  K t, s, u(s) ds,

(9.21)

a

where K is a continuous function of three variables K(t, s, u), twice differentiable with respect to u. Then  b,     T(x + h)(t) = K t, s, x(s) + Ku t, s, x(s) h(s) a

  2 1 + Kuu t, s, x(s) h (s) + · · · ds 2 1 = T(x) + T (x)h + T

(x)(h, h) + · · · , 2

(9.22)

where Ku = ∂K/∂u,



b

T (x)h =

  Ku t, s, x(s) h(s)ds

(9.23)

a

and



b

T (x)(h, h)(t) =

  Kuu t, s, x(s) h2 (s)ds.

(9.24)

a

The left-hand side of (9.24) represents the value at t of the function T

(x)(h, h) obtained by application of the bilinear operator T

(x) to the pair of functions (h, h). Clearly, T

(x) applied to a general pair of functions (h, k) yields  b  

Kuu t, s, x(s) h(s)k(s)ds. T (x)(h, k)(t) = (9.25) a

This can be obtained from (9.24) using (9.20).




488

Chapter 9

Miscellaneous Applications

Example 9.2.21. Consider the real Hilbert space L2 ([a, b]). Define a functional f : L2 ([a, b]) → R by the double integral f (x) =

 b

b

x(s)K(s, t)x(t)dtds, a

(9.26)

a

where K is a continuous function. If we define a linear operator T on L2 ([a, b]) by  b (Tx)(t) = K(t, s)x(s)ds, (9.27) a

we can rewrite (9.26) in a neat form: f (x) = x, Tx.

(9.28)

Hence, $ % f (x + h) = x + h, T(x + h) = x + h, Tx + Th = f (x) + h, Tx + x, Th + h, Th.

(9.29)

The second Fréchet differential f

is given by the quadratic term in (9.29), so that f

(x)(h, h) = h, Th.

(9.30)

Application of the formula (9.20) gives f

(x)(h, k) =

2 1   1 1 h, Tk + k, Th = h, T + T ∗ k , 2 2

(9.31)

where T ∗ is the adjoint of T. If K(s, t) = K(t, s), then T is self-adjoint, and (9.31) becomes f

(x)(h, k) = h, Tk.

(9.32)

This expression is symmetric in h and k. This result can be generalized to any functional f on a real Hilbert space.




Theorem 9.2.22. Suppose f : H → R is a twice differentiable functional on a real Hilbert space H. Then, for each x ∈ H, there exists a self-adjoint bounded linear operator T : H → H (depending on x) such that f

(x)(h, k) = h, Tk for all h, k ∈ H.

Proof: For any fixed k ∈ H, f

(x)(h, k) is a continuous linear functional. Continuity follows from Fréchet differentiability. According to the

9.2 The Gateaux and Fréchet Differentials

489

Riesz Representation Theorem 3.7.7, there exists an x0 ∈ H such that f

(x)(h, k) = h, x0 . Thus, there exists a map T which to each k ∈ H assigns x0 ∈ H, that is, x0 = Ak. Since f

(x)(h, k) is a linear in k for every fixed h, it follows that T is a linear operator. Consequently, f

(x)(h, k) = h, x0  = h, Tk. Moreover, since f

(x) is continuous and symmetric, T is bounded and selfadjoint.


Definition 9.2.23.

(Convex function) A function f : E → R, where E is a vector space, is called convex if   f tx + (1 − t)y ≤ tf (x) + (1 − t)f (y) for all x, y ∈ E, and t ∈ (0, 1). (9.33)

It is called strictly convex if   f tx + (1 − t)y < tf (x) + (1 − t)f (y)

for all x, y ∈ E, and t ∈ (0, 1). (9.34)

Example 9.2.24. One of the most important convex functions is the norm in a normed space: f (x) = x .




Example 9.2.25. Suppose S is a convex subset of a Hilbert space H. Define the distance function d(x, S) = inf x − y . y∈S

It follows from the triangle inequality that d(x, S) is a convex function.




Example 9.2.26. Let T be a self-adjoint bounded operator on a Hilbert space H. Consider the quadratic functional  defined by 1 (x) = Tx, x. 2  is differentiable and ∇ = T. Moreover,  is convex if T is strictly positive.


Theorem 9.2.27. If a convex function f : B → R is Gateaux differentiable at some

x0 ∈ B, then

f (x) ≥ f (x0 ) + df (x0 )(x − x0 )

for all x ∈ B.

490

Chapter 9

Miscellaneous Applications

Proof: Since f is convex, for every t ∈ (0, 1), we have

    f x0 + t(x − x0 ) − f (x0 ) ≤ tf (x) + (1 − t)f (x0 ) − f (x0 ) = t f (x) − f (x0 ) ,

and thus, f (x0 + t(x − x0 )) − f (x0 ) . t Now, by letting t → 0, we obtain f (x) − f (x0 ) ≥

f (x0 + t(x − x0 )) − f (x0 ) = df (x0 )(x − x0 ). t→0+ t

f (x) − f (x0 ) ≥ lim




9.3 Optimization Problems and the Euler–Lagrange Equations In ordinary calculus, maximum and minimum problems are concerned with those values of the independent variables for which a given function attains its maximum or minimum. If a differentiable function has a maximum or a minimum at a point, then its derivative vanishes at that point. It turns out that this property can be generalized to the case of a maximum or a minimum of a functional on a normed space. We will show that if a real-valued functional defined on a subset of a normed space has a maximum or a minimum at a point, then its Gateaux (or Fréchet) differential at that point is zero. In the subsequent discussion, we use the term extremum to refer to either maximum or minimum. Suppose f is a real-valued function defined on a subset  of a normed space. A very general formulation of an optimization problem is Find x0 ∈  such that f (x0 ) = min f (x). x∈

(9.35)

This and the following section are concerned with some aspects of such a problem.

Definition 9.3.1.

(Relative extremum) A real-valued functional f defined on a subset  of a normed space E is said to have a relative minimum (or relative maximum) at a point x0 ∈  if there is an open ball B(x0 , r) ⊂ E such that f (x0 ) ≤ f (x) (or f (x0 ) ≥ f (x)) holds for all x ∈ B(x0 , r) ∩ . If f has either a relative minimum or relative maximum at x0 , then f is said to have a relative extremum at x0 . A relative extremum is often called a local extremum. The set  on which an extremum problem is defined is sometimes called the admissible set.

9.3 Optimization Problems

491

Theorem 9.3.2. If a functional f : E → R is Gateaux differentiable at x0 ∈ E and has a relative extremum at x0 , then df (x0 , h) = 0 for all h ∈ E. Proof: For every h ∈ E, the function f (x0 + th) (of the real variable t) has a relative extremum at t = 0. Since it is differentiable at 0, it follows from ordinary calculus that ,

d f (x0 + th) dt

= 0. t=0

This means that df (x0 , h) = 0 for all h ∈ E, proving the theorem.




Corollary 9.3.3. If a functional f : E → R is Fréchet differentiable at x0 ∈ E and has a relative extremum at x0 , then f (x0 ) = 0.
Definition 9.3.4. (Stationary point) A point x at which df (x, h) = 0 for all h ∈ E (or f (x) = 0) is called a stationary point. The preceding theorem (or corollary) states that relative extrema of a Gateaux (or Fréchet) differentiable function occur at stationary points.

Example 9.3.5. Consider the real-valued functional f defined on a real Hilbert space H1 by f (x) = u − Tx 2 , where T is a bounded linear operator from H1 into a Hilbert space H2 , and u ∈ H2 . The Fréchet differential is given by f (x) = −2T ∗ u + 2T ∗ Tx. Therefore, a necessary condition for f to have an extremum at x0 is that f (x0 ) = 0, that is, T ∗ Tx0 = T ∗ u.




Theorem 9.3.6. Suppose f is a real-valued functional on a vector space E and x0 minimizes f on a convex set  ⊂ E. If f is a Gateaux differentiable at x0 , then df (x0 , x − x0 ) ≥ 0 for all x ∈ .

492

Chapter 9

Miscellaneous Applications

Proof: Since  is a convex set, x0 + t(x − x0 ) ∈  for all t ∈ (0, 1) and x ∈ . Hence,

,

 d  f x0 + t(x − x0 ) ≥ 0. dt t=0




Example 9.3.7. Consider a functional I : C ([0, 1]) → R defined by 

1

I(x) =

a(t)x2 (t)dt,

0

where a is a given function. Then  1  1 a(t)x(t)h(t)dt + a(t)h2 (t)dt. I(x + h) − I(x) = 2 0

Hence, I (x)h = 2

0



1

a(t)x(t)h(t)dt. 0

Consequently, I (x) = 0 if x = 0.




Example 9.3.8. Consider the functional I(u) = Au, u − 2u, f , where f belongs to a real Hilbert space H, and A is a linear self-adjoint operator on H. Clearly, I(u + h) − I(u) = 2Au, h − 2f , h + Ah, h. Setting I (u), h = 2Au, h − 2f , h = 2(Au − f ), h, we obtain   $ %  I(u + h) − I(u) − I (u), h  = Ah, h ≤ K h 2 , where K is a constant. Thus, it follows that the Fréchet differential of I(u) is I (u) = 2(Au − f ). Furthermore, it follows that % $ % $ I (u + h), k − I (u), k = 2Ah, k. Thus, the second Fréchet differential is I

(u)(h, k) = 2Ah, k which is independent of u ∈ H. It turns out that if A is a positive operator, then I(u) has a local minimum when u satisfies the operator equation Au = f .


9.3 Optimization Problems

493

The most remarkable classical Euler–Lagrange variational problem is to determine a function u(x) on the interval [a, b] satisfying the boundary conditions u(a) = α and u(b) = β, and extremizing the functional    b du

, (9.36) I(u) = F(x, u, u )dx u (x) = dx a where u is a twice continuously differentiable function on the interval [a, b] (u ∈ C 2 ([a, b])), F is continuous in x, u, and u , and has continuous partial derivatives with respect to u and u . We assume that I(u) has an extremum at some u ∈ C 2 ([a, b]). Then we consider the set of all variations u + tv, for an arbitrary fixed v ∈ C 2 ([a, b]), such that v(a) = v(b) = 0. Then  b   I(u + tv) − I(u) = (9.37) F(x, u + tv, u + tv ) − F(x, u, u ) dx. a

Using the Taylor series expansion   ∂F ∂F + v F(x, u + tv, u + tv ) = F(x, u, u ) + t v ∂u ∂u   ∂F 2 t 2 ∂F + v + v + · · · , 2! ∂u ∂u it follows from (9.37) that I(u + tv) = I(u) + tdI(u, v) +

t2 2 d I(u, v) + · · · , 2!

where the first and the second Fréchet differentials are given by   b ∂F

∂F dI(u, v) = v + v dx, ∂u ∂u a   b ∂F 2 ∂F + v dx. v d2 I(u, v) = ∂u ∂u a

(9.38)

(9.39) (9.40)

The necessary condition for the functional I to have an extremum at u is that dI(u, v) = 0 for all v ∈ C 2 ([a, b]) such that v(a) = v(b) = 0, that is,   b ∂F

∂F + v dx. v (9.41) 0 = dI(u, v) = ∂u ∂u a Integrating the second term in the integrand in (9.41) by parts, we obtain  ,  b, d ∂F ∂F ∂F b − vdx + v = 0. (9.42) ∂u dx ∂u ∂u a a

494

Chapter 9

Miscellaneous Applications

Since v(a) = v(b) = 0, the boundary terms vanish and the necessary condition becomes   b, d ∂F ∂F − vdx = 0 (9.43) ∂u dx ∂u a for all functions v ∈ C 2 ([a, b]) vanishing at a and b. This is possible only if   d ∂F ∂F − = 0. ∂u dx ∂u

(9.44)

This is called the Euler–Lagrange equation. We therefore have the following:

Theorem 9.3.9. (The Euler–Lagrange Variational Principle) A necessary condition for the functional I(u) to be stationary at u is that u must satisfy the Euler– Lagrange equation (9.44) in a ≤ x ≤ b with the boundary conditions u(a) = α and u(b) = β. After we have determined the solution of (9.44), which makes I(u) stationary, the question arises on whether I(u) has a minimum, a maximum, or a saddle point there. To answer this question, we look at the second derivative involved in (9.38). If terms of o(t 2 ) can be neglected in (9.38), or if they vanish for the case of quadratic F, it follows that a necessary condition for the functional I(u) to have a minimum at u0 is that d2 I(u, v) ≥ 0 for all v. Similarly, a necessary condition for the functional I(u) to have a maximum at u0 is that d2 I(u, v) ≤ 0 for all v. These results enable us to determine the upper and lower bounds for the stationary value I(u0 ) of the functional.

Example 9.3.10. (Minimum arc length) Determine the form of the curve in a plane, which will make the distance between two points in the plane minimum. Suppose the plane curve y = y(x) passes through the points (x1 , y1 ) and (x2 , y2 ). The length of such a curve is given by the functional  x2 I(y) = 1 + (y )2 dx. (9.45) x1

Thus, the problem is to determine the curve for which the functional I(y) is minimum. Since F = 1 + (y )2 depends on y only, the Euler–Lagrange equation becomes   d ∂F = 0. dx ∂y Hence, y

= 0.

(9.46)

9.3 Optimization Problems

495

This means that the curve extremizing I(y) is a straight line: y=

y2 − y1 (x − x1 ) + y1 . x2 − x1




Example 9.3.11. Determine the meridian curve joining two points in a plane which, when revolved about the x-axis, gives the surface of revolution with minimum area. This is a problem of minimum surface of revolution generated by the rotation of the curve y = y(x) about the x-axis. In this case, the area is given by  x2 S = 2π y(x) 1 + (y )2 dx, x1

so that the functional to be minimized is  x2 I(y) = y(x) 1 + (y )2 dx, x1

subject to the conditions y1 = y(x1 ) This corresponds to

and

y2 = y(x2 ).

(9.47)

F(x, y, y ) = y 1 + (y )2 ,

which does not depend on x explicitly. The Euler–Lagrange equation is yy

− (y )2 − 1 = 0.

(9.48)

Writing p for y , we have y

= dp/dx = pdp/dy, and (9.48) becomes py

dp = p2 + 1. dy

Separating the variables and integrating, we obtain y = a 1 + p2 , and hence dy = dx

.

y2 − 1. a2

Integrating again, we find  x−b , y = a cosh a 

(9.49)

496

Chapter 9

Miscellaneous Applications

where a and b are constants of integration, which can be determined from conditions (9.47). The curve defined by (9.49) is called the catenary, and the resulting surface is called a catenoid of revolution.


Example 9.3.12. Consider the functional I(u) =

 b, a

1 1 p(x)(u )2 − q(x)u2 + f (x)u dx, 2 2

(9.50)

where p, q, and f are given functions and u belongs to an admissible set  of I. Clearly, the Euler–Lagrange equation associated with the functional I(u) is   ∂F d ∂F = 0, (9.51) − ∂u dx ∂u where 1 1 F(x, u, u ) = p(x)(u )2 − q(x)u2 + f (x)u. 2 2 Consequently, (9.51) becomes (pu ) + qu = f . This is a nonhomogeneous ordinary differential equation of the Sturm– Liouville type.


Example 9.3.13. (Hamilton’s principle) According to Hamilton’s principle, a particle moves on a path that makes the time integral  t2 I= L(qi , q˙ i , t)dt

(9.52)

t1

stationary, where the Lagrangian L = T − V is the difference between the kinetic energy T and the potential energy V . In coordinate space, there are numerous possible paths joining any two positions. From all these paths, which start at a point A at time t1 and end at another point B at time t2 , nature selects the path qi = qi (t) for which dI = 0. Consequently, the Euler–Lagrange equation assumes the form   d ∂L ∂L − = 0, i = 1, 2, . . . , n. (9.53) ∂qi dt ∂ q˙ i In classical mechanics, these are simply called the Lagrange equations of motion. The Hamilton’s function (or Hamiltonian) H is defined in terms of the generalized coordinates qi , generalized momentum pi = ∂L/∂ q˙ i , and L as H=

n  i=1

pi q˙ i − L =

n  i=1

q˙ i

∂L − L(qi , q˙ i ). ∂ q˙ i

(9.54)

497

9.3 Optimization Problems

It follows that

 n    n  d ∂L d  ∂L dH ∂L = q˙ i −L = q˙ i − = 0. dt dt ∂ q˙ i dt ∂ q˙ i ∂qi i=1

i=1

Hence, the Hamiltonian H is the constant of motion.




Example 9.3.14. (Fermat’s principle in optics) This principle states that, in an optically homogeneous isotropic medium, light travels from one point (x1 , y1 ) to another point (x2 , y2 ) along a path y = y(x) for which the travel time is minimum. Since the velocity v is constant in such a medium, the time is minimum along the shortest path. In other words, the path y = y(x) minimizes the integral  x2  x2 1 + (y )2 I= F(y, y )dx, (9.55) dx = v x1 x1 with y(x1 ) = y1 and y(x2 ) = y2 . The Euler–Lagrange equation is given by   d

∂F F − y = 0. dx ∂y Hence, F − y

∂F = constant ∂y

or 1 1 = constant. v 1 + (y )2

(9.56)

To give a physical interpretation, we rewrite (9.56) in terms of the angle ϕ between the tangent to the minimum path and the vertical y-axis, so that sin ϕ =

1 1 + (y )2

.

Hence, sin ϕ = constant (9.57) v for all points on the minimum path. For a ray of light 1/v must be proportional to the refractive index n of the medium through which light is traveling. Equation (9.57) is known as the Snell law of refraction of light. This law is often stated as n sin ϕ = constant.

(9.58)


498

Chapter 9

Miscellaneous Applications

Example 9.3.15. (Abel’s problem of tautochronous motion) The problem is to determine the plane curve y = y(x) for which the time of descent of a particle sliding freely along the curve that passes through the origin and the point (x1 , y1 ) is minimum (see Example 5.7.1). The velocity of the particle at the intermediate point (x, y) is found from the energy equation  2 1 ds m = mg(y − 0), 2 dt so that the velocity is ds = 2gy, dt or



1 + (y )2 dx =



2gy dt,

which gives the time required for the particle to descend from the origin to the point (x1 , y1 ) on a frictionless curve path in a plane as  x1 + 1 + (y )2 T(y) = dx. 2gy 0 The problem is to minimize this functional subject to the conditions y(0) = 0, y(x1 ) = y1 . This case corresponds to + 1 + (y )2

. F(x, y, y ) = F(y, y ) = 2gy Thus, the Euler–Lagrange equation   ∂F d ∂F − =0 ∂y dx ∂y can be written as 0=y



d ∂F ∂F − ∂y dx ∂y



  d

∂F = F−y , dx ∂y

so that F − y

∂F = c, ∂y

where c is a constant. More explicitly, c 2gy(1 + (y )2 = 1,

499

9.3 Optimization Problems

or

+ y = ±

a−y , y

where a−1 = 2gc2 . This can be integrated to obtain  y ydy x= , ay − y2 0 where the positive sign is appropriate. We rewrite this integral in the form   1 y (a − 2y)dy a y dy x=− + , 2 2 0 2 ay − y ay − y2 0 so that the first integral can be evaluated at once, and the second one can be evaluated by making the substitution (a/2) − y = (a/2) cos θ . The final result is   a a − 2y x = − ay − y2 + cos−1 . 2 a This is the equation of the curve of minimum time of descent, where the constant a is to be determined so that the curve passes through the point (x1 , y1 ). It is convenient to write the equation in a parametric form by letting (a/2) − y = (a/2) cos θ . Then a x = (θ − sin θ ), 2

a y = (1 − cos θ ). 2

They represent a cycloid.




Example 9.3.16. A uniform elastic beam of length l is fixed at each end. The beam of line density ρ, cross-sectional moment of inertia I, and modulus of elasticity E, perform small transverse oscillations in the horizontal x-y-plane. Derive the equation of the motion of the beam. The potential energy of the beam is   1 l M2 1 l V= EI(y

)2 dx, (9.59) dx = 2 0 EI 2 0 where the bending moment M is proportional to the curvature so that M = EI

y

1 + (y )2

The kinetic energy is 1 T= 2

∼ EIy



l 0

ρ y˙ 2 dx.

for small y .

(9.60)

500

Chapter 9

Miscellaneous Applications

The variational principle gives 

t2

d



t2

Ldt = d

t1



t2

(T − V)dt = d

t1

F(˙y, y

)dt = 0,

(9.61)

t1

where



1 F(˙y, y ) = 2

l

 ρ y˙ 2 − EI(y

)2 dx.

(9.62)

0

The associated Euler–Lagrange equation is 

l

 ρ y¨ + EIy(iv) dx = 0

− 0

or ρ y¨ + EIy(iv) = 0.

(9.63)

This represents the equation of the motion for the transverse vibration of the beam.


Example 9.3.17. Find u(x, y, z) which minimizes the functional 





I u(x, y, z) = 

 2  ux + u2y + u2z dx dy dz,

where  ⊂ R3 . The Euler–Lagrange equation is       ∂ ∂F ∂ ∂F ∂ ∂F + + =0 ∂x ∂ux ∂y ∂uy ∂z ∂uz or uxx + uyy + uzz = 0, which is the three-dimensional Laplace equation.




We next consider optimization problems with a given auxiliary condition. It is often required to determine the function y = y(x) which minimizes the functional  x2 I(y) = F(x, y, y )dx (9.64) x1

subject to auxiliary condition  J(y) =

x2 x1

G(x, y, y )dx = C.

(9.65)

9.3 Optimization Problems

501

It can easily be shown that this problem is equivalent to the problem already discussed earlier, namely, that of determining the function y = y(x), which minimizes the functional  x2   I1 (y) = I(y) + λJ(y) = (9.66) F(x, y, y ) + λG(x, y, y ) dx. x1

The constant λ involved in (9.66) must be determined from the auxiliary condition (9.65). The resulting Euler–Lagrange equation is , d ∂ ∂ (F + λG) − (F + λG) = 0, ∂y dx ∂y

(9.67)

which must be solved with the auxiliary condition (9.65). However, this method breaks down in the case where   d ∂G ∂G = 0. −

dx ∂y ∂y

Example 9.3.18. The problem is to find the curve y = y(x) of the shortest length between two points (x1 , y1 ) and (x2 , y2 ) such that the area under the curve is A. The length of the curve is given by the functional  x2 (9.68) 1 + (y )2 dx I(y) = x1

and the area under the curve is 

x2

y(x)dx = A.

(9.69)

x1

This is a constrained optimization problem which reduces to that of finding the extremum of  x2   1 + (y )2 + λy dx. (9.70) I1 (y) = x1

The associated Euler–Lagrange equation is ,   d ∂  ∂ 

2

2 1 + (y ) + λy = 0 1 + (y ) + λy − ∂y dx ∂y or d y = λ. dx 1 + (y )2

(9.71)

502

Chapter 9

Miscellaneous Applications

This differential equation can be integrated twice to obtain the equation for y: 

α x− λ

2 + (y − β)2 =

1 . λ2

(9.72)

Thus, the curve of shortest length is an arc of a circle, where the constants of integration α and β together with the constant λ can be determined from the condition that the curve passes through the points (x1 , y1 ) and (x2 , y2 ) and the given constrained condition (9.69).


Example 9.3.19. (Geodesic) A geodesic is a curve of minimum length between two points on a smooth surface G(x, y, z) = 0 when the whole curve is confined to the surface. The geodesic of the surface G = 0 is obtained by minimizing the functional  I(x, y, z) =

t2



x˙ 2 + y˙ 2 + z˙ 2 dt,

(9.73)

t1

subject to the auxiliary condition G(x, y, z) = 0. This extremum problem is equivalent to finding the minimum of the functional  t2   I1 (x, y, z) = x˙ 2 + y˙ 2 + z˙ 2 + λ(t)G(x, y, z) dt. (9.74) t1

A necessary condition for this minimum problem leads to the Euler– Lagrange equations: d x˙ + λ(t)Gx = 0, dt x˙ 2 + y˙ 2 + z˙ 2

(9.75a)

y˙ d + λ(t)Gy = 0, dt x˙ 2 + y˙ 2 + z˙ 2

(9.75b)

d z˙ + λ(t)Gz = 0. dt x˙ 2 + y˙ 2 + z˙ 2

(9.75c)

These equations, combined with the constraint condition G(x, y, z) = 0, can be solved for the desired geodesic x(t), y(t), z(t), and λ(t). In particular, the geodesic on a sphere G(x, y, z) = x2 + y2 + z2 − r2 = 0 can be determined. In this case, Gx = 2x, Gy = 2y, and Gz = 2z. Let ax + by + cz = 0

9.3 Optimization Problems

503

be an equation of the plane which contains two given points on the sphere and the origin. Let the curve of minimum length be given by p(t) = ax(t) + by(t) + cz(t). Clearly, p(t) satisfies the equation d p˙ (t) + 2λ(t)p(t) = 0, 2 dt x˙ + y˙ 2 + z˙ 2 and p(t1 ) = p(t2 ) = 0. It follows from the uniqueness of solutions of differential equations that p(t) = 0. Therefore, the geodesic lies on the plane and is a segment of a great circle of the sphere.
The variational problem can readily be extended for functionals depending on functions of several variables. Many physical problems require us to determine a function of several variables which will give rise to an extremum of such functionals. The following two-variable case will serve as an illustration of such a problem. The associated functional has the form    I u(x, y) = F(x, y, u, ux , uy )dx dy, (9.76) 

where u is a function of two real variables. We want to find the function u that extremizes I. The values of u are prescribed on the boundary ∂ of . The function F is defined on the domain  of x and y, and is assumed to have continuous second-order partial derivatives. In accordance with the preceding case of functionals depending on a function of one independent variable, the first variation δI of I is defined by δI(u, ε) = I(u + ε) − I(u). By Taylor’s expansion theorem, this reduces to  δI = (εFu + εx Fp + εy Fq )dx dy,

(9.77)



where ε = ε(x, y) is small, p = ux , and q = uy . A necessary condition for the functional I to have an extremum is that the first variation of I vanishes, that is,  0 = δI = (εFu + εx Fp + εy Fq )dx dy  =



 ∂ ∂ ε Fu − Fp − Fq dx dy ∂x ∂y  

504

Chapter 9

Miscellaneous Applications   ,  ∂ ∂ Fp + Fq + (εx Fp + εy Fq ) dx dy + ε ∂x ∂y     ∂ ∂ = ε Fu − Fp − Fq dx dy ∂x ∂y   , ∂ ∂ + (εFp ) + (εFq ) dx dy. ∂y  ∂x

(9.78)

We assume that the boundary curve ∂ has a piecewise continuously turning tangent so that Green’s theorem can be applied to the second double integral in (9.78). Consequently, (9.78) becomes     ∂ ∂ 0 = δI = ε Fu − Fp − Fq dx dy + ε(Fp dy − Fq dx). (9.79) ∂x ∂y  ∂ Since ε is assumed to vanish on the curve ∂, the second integral in (9.79) equals 0. Moreover, since ε is otherwise arbitrary, it follows that the integrand of the first integral in (9.79) must vanish identically. Consequently, the extremizing function u(x, y) satisfies Fu −

∂ ∂ Fp − Fq = 0. ∂x ∂y

(9.80)

This is the Euler–Lagrange partial differential equation for the variational problem involving two independent variables. A necessary condition for an extremum of I[u(x, y)] is that u satisfies this equation. If the values of u are prescribed on the boundary ∂, the solution of the Euler–Lagrange equation which satisfies the given boundary conditions must be found.

Example 9.3.20. (Vibrating string) We derive the equation of motion for free vibration of an elastic string of length l and line density ρ. We assume that, initially, the string is stretched along the x-axis from x = 0 to x = l. The string will be given a small lateral displacement, which at each point along the x-axis is denoted by u(x, t). The kinetic energy of the string is given by  1 l 2 T= ρut dx, (9.81) 2 0 provided longitudinal motion of the string is neglected. If the displacement of the string is small, the change in length of the string caused by the displacement is given by  l 0

1 + u2x dx − l

  l  1 2 1 l 2 1 + ux dx − l = ∼ u dx. 2 2 0 x 0

9.4 Minimization of Quadratic Functionals

505

Since the potential energy is proportional to the elongation, we have V=

T1 2



l

0

u2x dx,

(9.82)

where T1 is the constant tension of the string. According to the Hamilton principle,  0=d

t2

t1

 Ldt = d

t2

 (T − V)dt = d

t1

t2 t1

1 2



l 0

 ρu2t − T1 u2x dx dt.

(9.83)

In this case, L = 12 (ρu2t − T1 u2x ), which does not depend explicitly on x, t, or u, and hence, the above result has the form 

t2



l

d t1

L(ut , ux )dx dt = 0.

0

Thus, the Euler–Lagrange equation is ∂ ∂ (ρut ) − (T1 ux ) = 0 ∂t ∂x or utt − c2 uxx = 0,

(9.84)

where c2 = T1 /ρ. This is the familiar partial differential equation for the vibrating string.


9.4 Minimization of Quadratic Functionals Suppose A is a real symmetric positive definite operator defined on a Hilbert space H. We study the variational formulation of the operator equation Au = f ,

(9.85)

where f is a given element of H. The solution of (9.85), if it exists, is unique due to the fact that the operator A is positive definite. Moreover, the solution of (9.85) can be shown to be the function that minimizes the quadratic functional I(u) = Au, u − 2f , u.

(9.86)

Conversely, if we can find a solution u that minimizes I(u) on H, then u is the desired solution of (9.85). This fundamental result is stated in the following theorem.

506

Chapter 9

Miscellaneous Applications

Theorem 9.4.1. Suppose A : H → H is a linear symmetric positive definite operator on a real Hilbert space H and f is a given element of H. Then the quadratic functional I(u) = Au, u − 2f , u attains its minimum value for some u0 ∈ H if and only if u0 is the solution of the equation Au = f .

Proof: Suppose u0 is the solution of (9.85). Let u be any element of H. Then I(u) − I(u0 ) = Au, u − 2f , u − Au0 , u0  + 2f , u0  = Au, u − 2Au0 , u + Au0 , u0  $ % = A(u − u0 ), u − u0 . Since A is a positive definite operator, A(u − u0 ), u − u0  ≥ 0, where the equality holds if and only if u = u0 . Hence, I(u) ≥ I(u0 ). This shows that I(u) attains its minimum at the solution u0 of (9.85). Conversely, suppose u0 is an element of H which minimizes I(u), that is, I(u) ≥ I(u0 ) for all u ∈ H. In particular, I(u0 + tv) ≥ I(u0 ) for any real number t and any v ∈ H. More explicitly, $ % I(u0 + tv) = A(u0 + tv), u0 + tv − 2f , u0 + tv = Au0 , u0  + 2tAu0 , v + t 2 Av, v − 2f , u0  − 2tf , v, or I(u0 + tv) − I(u0 ) = 2Au0 − f , v + tAv, v. t In the limit t → 0, this expression leads to the Gateaux differential dI(u0 , v) = 2Au0 − f , v for any v ∈ H. Since I(u) has a local minimum at u = u0 , dI(u0 , v) = 0 for all v ∈ H. This implies that Au0 − f = 0, which proves that u0 is a solution of (9.85).
This minimization problem can be interpreted as a maximization problem for −I(u). Theorem 9.4.1 can be generalized to a symmetric positive definite operator A (not necessarily real) on a complex Hilbert space H. The functional I is then modified to I(u) = Au, u − f , u − u, f .

9.5 Variational Inequalities

507

Note that Theorem 9.4.1 can be interpreted physically as a minimum energy principle.

Example 9.4.2. Suppose A : H → K is a symmetric bounded linear operator, where H and K are real Hilbert spaces. We want to minimize I(u) = Au − b 2 , where u ∈ H and b ∈ K. We have I(u) = Au − b, Au − b = Au, Au − 2b, Au + b, b % $ % $ = A∗ Au, u − 2 A∗ b, u + b, b, and thus, % $ % $ I(u + h) − I(u) = 2A∗ Au − 2A∗ b, h + A∗ Ah, h . Clearly, I(u) has the first differential given by I (u) = 2A∗ Au − 2A∗ b. Thus,

% $ % $ % $ I (u + h), k − I (u), k = 2A∗ Ah, k .

Hence, the second differential is $ % I

(u)(h, k) = 2A∗ Ah, k , which is independent of u. So I has an extremum at u = u0 if I (u0 ) = 0, that is, A∗ Au0 = A∗ b. In particular, if A∗ A is a positive operator, then I

(u) ≥ 0, and u0 given by the above equation is a local minimum of I(u).


9.5 Variational Inequalities Most partial differential equations in physical and engineering sciences arise from a variational principle. Usually, there is a class of admissible solutions and an energy functional associated with these admissible functions. We seek to minimize the energy to determine the solution of the problem which satisfies the Euler–Lagrange equation. Many physical problems can be expressed in terms of an unknown function u, for example, displacement of a mechanical system, satisfying an inequality a(u, v − u) ≥ F(v − u)

for all v ∈ S,

(9.87)

508

Chapter 9

Miscellaneous Applications

where the set S of admissible functions is a closed convex subset of a Hilbert space H, a(·, ·) is a bilinear form, and F is a bounded linear functional on S. Such inequalities are called variational inequalities. So the existence and uniqueness of the solutions of this type of inequality are of special interest. Under suitable conditions, the problem of the variational inequality is found to be equivalent to the minimization problem. Find an element u such that u∈S

and I(u) = inf I(v), v∈S

where the functional I : S → R is defined by 1 I(v) = a(v, v) − F(v). 2

(9.88)

The problem of finding u ∈ S is called an abstract minimization problem. The problem of finding u ∈ S such that a(u, v − u) ≥ F(v − u)

(9.89)

for all v ∈ S is called a variational inequality problem, and u is called its solution. The following theorem gives the relation between the abstract minimization problem and the variational inequality problem.

Theorem 9.5.1. Suppose a(·, ·) is a continuous, symmetric, and elliptic bilinear form on a Hilbert space H, f ∈ H, and S ⊂ H is a closed convex subset. Then there exists a unique u ∈ S such that a(u, v − u) ≥ f , v − u

for each v ∈ S.

(9.90)

Further, u can be characterized by u ∈ S,

(9.91a)

I(u) = min I(v), v∈S

(9.91b)

where 1 I(v) = a(v, v) − f , v. (9.92) 2 Proof: Since a(u, v) is a symmetric and elliptic bilinear form on H, it defines an inner product ·, ·a and a norm · a on H: u, va = a(u, v), u a = a(u, u).

(9.93a) (9.93b)

By the continuity and ellipticity of a(u, v), we have K u 2 ≤ a(u, v) ≤ M u 2 ,

(9.94)

9.5 Variational Inequalities

509

and so the norm · a is equivalent to the original norm · on H. Consequently, H is a Hilbert space with respect to the new inner product. Now, by the Riesz Representation Theorem 3.7.7, there exists fˆ ∈ H such that, for every v ∈ H,   a fˆ , v = f , v. (9.95) Now,    1 v − fˆ 2 = 1 a v − fˆ , v − fˆ a 2 2   1   1 = a(v, v) − a v, fˆ + a fˆ , fˆ 2 2 2 1 1 = a(v, v) − f , v + ˆf a (by (9.95)) 2 2 2 1 = I(v) + ˆf a . 2 Since fˆ 2a is a constant, minimizing I(v) over S is equivalent to minimizing v − fˆ 2a over S. So, by Theorem 3.6.5, there exists a unique u ∈ S such that % $ fˆ − u, v − u ≤ 0 for every v ∈ S, a

or

$

% fˆ , v − u a ≤ u, u − va

for every v ∈ S,

or, finally, a(u, u − v) ≥ f , v − u for every v ∈ S.




In the following theorem, we prove existence and uniqueness of the solution of the variational inequality (9.90) without assuming that the form a(·, ·) is symmetric.

Theorem 9.5.2. (Lions–Stampacchia) Suppose H is a Hilbert space, a(·, ·) is a continuous and elliptic bilinear form on H, and S is a closed convex subset of H. Then the variational inequality (9.90) has a unique solution.

Proof: For a fixed u ∈ H, we consider a map v → a(u, v). By the continuity of a(·, ·), this is a continuous linear functional. Hence, there exists Au ∈ H such that Au, v = a(u, v)

for every v ∈ H.

(9.96)

Since a(u, v) is bilinear, the map u → Au is linear. Further, it follows from continuity of a(u, v) that Au ≤ M u

510

Chapter 9

Miscellaneous Applications

for some M > 0. Hence A is continuous. Moreover, since a(u, v) is elliptic, we have Au, u ≥ K u 2 for some K > 0. With this new notation, we seek u ∈ S such that Au, v − u ≥ f , v − u for every v ∈ S.

(9.97)

Let α > 0 be a positive constant to be chosen shortly. Then (9.97) is equivalent to αf − αAu + u − u, v − u ≤ 0 for every v ∈ S. In other words, by Theorem 3.6.5, we seek u ∈ S such that u = P(αf − αAu + u), where P is the projection on S. Hence, we look for a fixed point of the continuous map F : H → H, whose range lies in S, defined by F(v) = P(αf − αAv + v). Now, if v1 , v2 ∈ H, we have     F(v1 ) − F(v2 )2 = P(αf − αAv1 + v1 ) − P(αf − αAv2 + v2 )2  2 ≤ (v1 − v2 ) − αA(v1 − v2 ) . Thus,

  $ % F(v1 ) − F(v2 )2 ≤ v1 − v2 2 − 2α a(v1 − v2 ), (v1 − v2 )  2 + α 2 A(v1 − v2 )   ≤ 1 − 2αK + α 2 M 2 v1 − v2 2 .

We next choose α such that 0 < α < 2K/M 2 ; then β 2 = 1 − 2αK + α 2 M 2 < 1. Clearly,   F(v1 ) − F(v2 ) ≤ β v1 − v2 with 0 < β < 1, and so F is a contraction. By the Banach Fixed Point Theorem (Theorem 1.6.4), F has a unique fixed point u which must belong to S.


9.6 Optimal Control Problems for Dynamical Systems The theory of optimal control is one of the major areas of applications of Hilbert space methods. A large class of optimal control problems can be for-

9.6 Optimal Control Problems

511

mulated as variational or optimizational problems. Indeed, optimal control problems are a special kind of optimization problem where the constraints specify the dynamics of the system described by differential equations. A control problem involves, first of all, a dynamical system — an ordinary differential equation with time as the independent variable:   dx = f t, x(t), u(t) , dt

0 ≤ t ≤ T,

x(0) = x0 ,

(9.98)

where x(t), a function with a range in a Euclidean space, is called the state function, and u(t) is referred to as the control function and is assumed to be in some specified set of admissible controls. The control problem is to find a control function u(t) which minimizes an “index of performance” or “cost functional”  T   I(u) = l(x, u, t)dt + m x(T) , (9.99) 0

subject to constraints on the control such as u(t) belongs to a convex set , and x(T) is in a given set , where l and m are given functions. The control problem is called time invariant if f and l are independent of t. The control problem is called linear-quadratic if the cost functional I(u) is quadratic, that is, l and m are quadratic in x and u, and the differential equation (9.98) is linear. A control that attains the minimum of performance index in the admissible class of controls is known as optimal control, and the corresponding x(t) as an optimal trajectory. The upshot of the theory is that under certain conditions an optimal control must satisfy the maximum principle of Pontrjagin. Suppose I = (t1 , t2 ) ⊂ R is an interval, A = [αij ] is a constant N × N matrix, and A(t) = [aij (t)] is an N × N matrix with elements aij (t), which are continuous functions on the interval I. Let xT = (x1 , . . . , xN ) be an N dimensional vector, and let vT (t) = (v1 (t), . . . , vN (t)) be an N dimensional vector with components vk (t), which are piecewise continuous on I. We put
  D = (t, x): t ∈ I, x ∈ RN or CN . With this notation, we consider linear systems of first order ordinary differential equations dx = A(t)x + v(t), dt dx = A(t)x, dt and

dx = Ax. dt

(9.100) (9.101)

(9.102)

512

Chapter 9

Miscellaneous Applications

According to the general theory of ordinary differential equations, systems (9.100), (9.101), and (9.102) possess unique solutions for every (t, x) ∈ D that exist over the entire interval I and that depend continuously on the initial data. A set on N linearly independent solutions of (9.101) on I is called the fundamental set of solutions. An N × N matrix  whose N columns are linearly independent solutions of (9.101) on I is called a fundamental matrix. If {1 , . . . , N } is a set of N linearly independent solutions of (9.101) and if (k )T = {1k , . . . , Nk }, then 

 1N 2N  ..  . 

11  21 =  ...

12 22 .. .

... ... .. .

N1

N2

. . . NN

is a fundamental matrix. We use the natural basis for the x-space:   1 0 ... 0 0 1 ... 0 . [u1 , . . . , uN ] =  .. ..   ... ... . . 0 0 ... 1 A fundamental matrix  for equation (9.101), whose columns are determined by the linearly independent solutions (τ ) = uk , k = 1, . . . , N, τ ∈ I, is called the state transition matrix  of equation (9.101). We now formulate a linear-quadratic optimal problem described by the dynamical equation for the state as dx = Ax(t) + Bu(t), dt

x(0) = x0 ,

(9.103)

where x(t) ∈ RN is the state vector, u(t) ∈ RM is the control vector, t ∈ [0, T], and B is an N × M matrix. According to the theory of ordinary differential equations, if every element of the vector u(t) is a continuous function of t, then the unique solution of (9.103) is given by 

t

x(t) = (t, 0)x(0) +

(t, τ )Bu(τ )dτ,

(9.104)

0

where (t, τ ) is the state transition matrix for the dynamical system (9.103). We next define the class of vector valued functions  
   L2M [0, T] = u: uT = (u1 , . . . , uM ), uk ∈ L2 [0, T] , k = 1, . . . , M .

9.6 Optimal Control Problems

513

Clearly, L2M ([0, T]) is a vector space. We introduce an inner product in L2M ([0, T]) by  T uT (t)v(t)dt. (9.105) u, v = 0

L2M ([0, T])

Since is complete, it is a Hilbert space (see Exercise 8). Next we define a linear operator L : L2M ([0, T]) → L2N ([0, T]) by  t [Lu](t) = (t, τ )Bu(τ )dτ. (9.106) 0

Since the elements of (t, τ ) are continuous functions on [0, T] × [0, T], L is a compact operator. The linear-quadratic control problem is to find u(t) ∈ L2M ([0, T]), which minimizes the cost functional  T  T T I(u) = x (t)x(t)dt + α uT (t)u(t)dt, (9.107) 0

0

where α is a positive constant which, without loss of generality, can be taken to be 1. If we replace −(t, 0)x0 by v(t) in (9.104) and rewrite the resulting equation in terms of the operator L, we obtain x(t) = Lu − v. Thus, the control problem reduces to a minimization problem in a Hilbert space with the cost functional I(u) = Lu − v, Lu − v + αu, u.

(9.108)

The required minimizing u(t) can be determined from the following general result.

Theorem 9.6.1. Suppose H and K are real Hilbert spaces and L : H → K is a

compact operator with L∗ as its adjoint. Let v be a given fixed element in K and let α ∈ R. Define a functional I : H → R by I(u) = Lu − v 2 + α u 2 ,

(9.109)

where · is the norm in K. If α > 0, then there exists a unique u0 ∈ H such that I(u0 ) ≤ I(u) for all u ∈ H. Moreover, u0 is a solution of the equation L∗ Lu0 + αu0 = L∗ v.

(9.110)

Proof: Let A = L∗ L. Then A is a positive compact operator. Therefore, −α

cannot be an eigenvalue of L∗ L. It can be shown that the equation Ax −λx = y has a unique solution provided α is not an eigenvalue of A and α = 0. The solution is given by ∞

x=

P0 y  Pn y + , −λ n=1 λn − λ

(9.111)

514

Chapter 9

Miscellaneous Applications

where λ1 , λ2 , . . . are nonzero distinct eigenvalues of A, Pn is the projection onto the null space of A − λn I , and P0 is the projection onto the null space of A. Suppose u0 ∈ H is the unique solution of (9.110). It follows from (9.109) that, for an arbitrary h ∈ H, we have I(u0 + h) = Lu0 + Lh − v, Lu0 + Lh − v + αu0 + h, u0 + h = Lu0 − v, Lu0 − v + 2Lh, Lh0 − v + v, v + αu0 , u0  + 2αu0 , h + αh, h = Lu0 − v, Lu0 − v + v, v + αu0 , u0  $ % + 2α h, L∗ Lu0 + αu0 − L∗ v + αh, h = Lu0 − v 2 + v 2 + α u0 2 + α h 2 . Hence, I(u0 + h) is minimum if and only if h = 0.

(9.112)


The solution of (9.110) can be obtained from (9.111). However, a more convenient technique is available for determination of the solution when L is given by (9.106). The method is formulated in the following theorem.

Theorem 9.6.2. Suppose I(u) is given by 

T

I(u) =



T

x (t)x(t)dt + α T

0

uT (t)u(t)dt,

0

where α > 0. Suppose x(t) satisfies the dynamical system (9.103): dx = Ax(t) + Bu(t), dt

x(0) = x0 .

If the control vector 1 (9.113) u(t) = − BT P(t)x(t), α for all t ∈ [0, T], and P(t) is the solution of the matrix Riccati equation ˙ = AT P(t) + P(t)A − P(t)

1 P(t)BBT P(t) + I = 0 α

(9.114)

with P(T) = 0, then u(t) minimizes I(u).

Proof: We will show that u(t) satisfies (9.110), where Lu is given by (9.106). If u(t) satisfies (9.110), then 1 1 u = − L∗ (Lu − v) = − L∗ x. α α

9.6 Optimal Control Problems

515

We next find a formula for evaluating L∗ w for an arbitrary w ∈ L2M ([0, T]). To do this, we calculate -T  T , s Lu, w = (s, t)Bu(t)dt w(s)ds 

0

= 

0

s

uT (t)BT T (s, t)w(s)dtds

0 T

=

0

T

,

T

T

u (t) 0

$ % B  (s, t)w(s)ds dt = u, L∗ w . (9.115) T

T

t

Consequently, 



∗



T

L w (t) =

Bt t (s, t)w(s)ds

t

for all t ∈ [0, T]. We next assume that there exists a matrix P(t) such that  T T (s, t)x(s)ds P(t)x(t) =

(9.116)

t

with P(T) = 0. It is necessary to find conditions for existence of such a matrix P(t). Differentiating (9.116) with respect to t and using the equality P˙ T (s, t) = −AT (s, t), we obtain ˙ P(t)x(t) + P(t)˙x(t) = −x(t) − AT



T

T (s, t)x(s)ds

t

= −x(t) − AT P(t)x(t).

(9.117)

In view of (9.103), this equation becomes   ˙ P(t)x(t) + P(t) Ax(t) + Bu(t) = −x(t) − At P(t)x(t). We next use (9.113) to replace u(t), so that the above equation becomes ˙ P(t)x(t) + P(t)Ax(t) −

1 P(t)BBT P(t)x(t) + x(t)I + AT P(t)x(t) = 0. α

Clearly, P(t) satisfies (9.114) with P(T) = 0. If 1 u(t) = − BT P(t)x(t), α

516

Chapter 9

Miscellaneous Applications

it turns out that u(t) satisfies L∗ Lu + αu = L∗ v, where v = −(t, 0)x0 , and hence by Theorem 9.6.1, u(t) minimizes the cost functional I(u).
The matrix Riccati equation (9.114) is often called the state equation of the linear-quadratic control problem, and it can be shown that this equation has a unique solution for all t < T.

Example 9.6.3. Find the control u which minimizes the cost functional 

T

I(u) =

 x2 + u2 dt,

0

where x(t) satisfies x˙ (t) = u(t) with x(0) = x0 . In this problem, A = 0, B = 1, and α = 1. Hence, (9.114) becomes ˙ − P2 (t) + 1 = 0, P(t)

P(T) = 0.

The solution of this equation is     P(t) = 1 − e2(t−T) / 1 + e2(t−T) . Thus, this optimal control u(t) is obtained from (9.113) in the form u(t) = −P(t)x(t), where x(t) can be solved from the given equation, x˙ (t) = u(t).


Example 9.6.4. (Control of chemical reaction with nonlinear cost) Consider a chemical mixture which is added to a tank at a constant rate for a fixed time interval [0, T]. Suppose that the pH value x at which the reaction occurs determines the quality of the final product and that this pH value can be controlled by the strength u(t) of some component of the mixture. Suppose the chemical reaction takes place so that the rate of change of pH value x(t) is proportional to the sum of the current pH value and the strength u(t) of the controlling component, that is, x˙ = ax + bu, where a and b are known positive constants. Assume that the decrease in yield T due to variation of the pH value is 0 x2 dt and the rate of cost of maintaining the strength u(t) is proportional to u2 . Then the total cost associated with the control function u(t) in [0, T] is  T  2  I(u) = cx + u2 dt, 0

where c > 0 is an appropriate constant. Suppose the initial pH is given, that is, x(0) = x0 . We seek a control function u∗ (t) on [0, T], which determines a

517

9.7 Approximation Theory

response x∗ (t) so that the cost functional is minimum. This is an example of a linear-quadratic optimal control problem.


9.7 Approximation Theory Typical problems in approximation theory deal with the determination of an element from a given set of elements that is closest to a prescribed element not in the given set. Such a closest point may or may not exist in general. More precisely, given an element x and a set S in a normed linear space E, an element y of S is said to be a best approximation to x from S if, for all z ∈ S, x − y ≤ x − z .

(9.118)

The problem of determining such an element y is usually called a best approximation problem. One of the most common and remarkable examples of such a problem is the approximation of continuous functions by polynomials. In the language of the optimization theory, the closest element y is called an optimal solution of the problem: Minimize x − z , where x ∈ E and z ∈ S ⊂ E. The element x − y is called an optimal error. The theory of approximation is essentially concerned with the following four basic problems: (a) Existence of best approximations. (b) Uniqueness of best approximations. (c) Characterization of best approximations. (d) Methods for determining the best approximation. A basic existence Theorem 3.6.4 was proved in Section 3.6. Of primary interest here are the uniqueness and characterization of best approximations.

Theorem 9.7.1. (Uniqueness of best approximation) Let S be a subspace of a Hilbert space H, and x ∈ H. If a best approximation of x from S exists, then it is unique. A necessary and sufficient condition that y ∈ S be the unique best approximation to x from S is that the optimal error x − y be orthogonal to S. Proof: Let y ∈ S be a best approximation to x from S. To prove (x − y) ⊥ S, we consider z ∈ S with z = 1. Then w = y + x − y, zz ∈ S. We have x − y 2 ≤ x − w 2 = x − w, x − w

(9.119)

518

Chapter 9

Miscellaneous Applications

$ % = (x − y) − x − y, zz, (x − y) − x − y, zz  2 = x − y, x − y − x − y, z  2 = x − y 2 − x − y, z . Consequently,

(9.120)

  x − y, z2 = 0,

that is, x − y, z = 0 or (x − y) ⊥ z. Conversely, let y ∈ S and (x − y) ⊥ S. Then, for any z ∈ S, we have y = z ∈ S so that (x − y) ⊥ (y − z). Hence, by Pythagorean Formula (3.8), x − z 2 = x − y + y − z 2 = x − y 2 + y − z 2 . Thus, x − z > x − y

if y = z.

This shows that y is the unique best approximation to x from S.




Note that the element y is the projection of x onto S. If S is a plane through the origin in R3 , the previous theorem states that the optimal error is obtained by dropping a perpendicular from x onto S. One of the most important problems in approximation theory is the leastsquares polynomial approximation in an interval. We first give the characterization of least-squares approximation and then demonstrate the unique importance of orthogonal polynomials in approximation theory. Let n denote the space of all polynomials of a single variable of degree at most n. It is a consequence of the basic existence theorem (Theorem 3.6.4) that, given a continuous function f (x) on the interval I = [−1, 1] and an integrable weight function w(x), which is positive on I, except possibly at a finite number of points of I at which w(x) = 0, there exists a unique polynomial Q∗n (x) ∈ n such that , 1 -1/2   2  ∗  f − Q∗  = < f − P 2 (9.121) f (x) − Qn (x) w(x)dx n 2 −1

for any P ∈ n , P = Q∗n . This Q∗n is called the least-squares approximation to f out of n , with respect to the weight function w(x). The following theorem characterizes the least-squares approximation.

Theorem 9.7.2. If f ∈ C (I), and w is a weight function in I, then Q∗n is the leastsquares approximation to f out of n if and only if  1   f (x) − Q∗n (x) P(x)w(x)dx = 0 −1

(9.122)

9.7 Approximation Theory

519

for every P ∈ n .

Proof: It suffices to use Theorem 9.7.1 with S = n and H defined as the space of functions such that  1 −1

  f (x)2 w(x)dx < ∞

with the inner product  f , g =

1

−1

f (x)g(x)w(x)dx.

(9.123)


Q∗n

To find explicitly, we put P(x) = x , k = 0, 1, 2, . . . , n, successively in (9.122) to obtain  1  1 xk Q∗n (x)w(x)dx = xk f (x)w(x)dx. (9.124) k

−1

−1

This is a system of (n+1) linear equations for the (n+1) unknown coefficients Q∗n (x). Suppose we write Q∗n (x) = a0 + a1 x + · · · + an xn ;

(9.125)

then the system (9.124) may be expressed as n 

ckm am = bk ,

k = 0, 1, 2, . . . , n,

(9.126)

m=1

where

 ckm =  bk =

1 −1 1 −1

xk+m w(x)dx,

(9.127)

xk f (x)w(x)dx.

(9.128)

In principle, the system of normal equations (9.126) can be solved to determine a0 , a1 , . . . , an , and hence Q∗n is determined explicitly. However, when n is large, say n ≥ 6, there are formidable numerical difficulties in solving the normal system even in the simple case w(x) = 1. These computational difficulties can be avoided by noting that {1, x, x2 , . . . , xn } is not the only set of polynomials that spans n . If P0 , P1 , . . . , Pn ∈ n are linearly independent, then every P ∈ n has a unique representation of the form P(x) = C0 P0 + C1 P1 + · · · + Cn Pn ,

(9.129)

520

Chapter 9

Miscellaneous Applications

where the set {P0 , P1 , . . . , Pn } can be determined by assuming that this set is orthogonal in −1 ≤ x ≤ 1 with respect to the given weight function w(x). In other words, the condition of orthogonality with respect to the inner product (9.123) requires  1 Pm (x)Ps (x)w(x)dx = δms , (9.130) −1

where m, s = 0, 1, 2, . . . , n. It is important to point out that {P0 , P1 , . . . , Pn } is a set of orthonormal polynomials with respect to w(x), and such an orthonormal set can always be constructed by the Gram–Schmidt orthonormalization process. One of the remarkable features of the orthonormal set is the ability to simplify the leastsquares approximation problem. This can be shown as follows: Suppose the set P0 , P1 , . . . , Pn ∈ n satisfies (9.130). Then the set is linearly independent. We next suppose Q∗n (x) =

n 

λk Pk .

(9.131)

k=0

We substitute P = Pm , m = 0, 1, . . . , n, successively in (9.122) to obtain  1  1 Pm (x)Q∗n (x)w(x)dx = Pm (x)f (x)w(x)dx. −1

−1

In view of (9.131), this result reduces to   1 n  λk Pm (x)Pk (x)w(x)dx =

1

Pm (x)f (x)w(x)dx

(9.132)

which, together with (9.130), determines λm ’s:  1 Pm (x)f (x)w(x)dx, m = 0, 1, . . . , n. λm =

(9.133)

k=0

−1

−1

−1

The system of normal equations is completely uncoupled, and the computational difficulties in solving the normal system (9.126) disappear as we choose an orthogonal basis for n . To obtain a set of polynomials orthonormal in −1 ≤ x ≤ 1 with respect to w(x), we define 1, x . (9.134) P˜ 0 (x) = 1, P˜ 1 (x) = x − 1, 1 Note that 1, 1 > 0 and P˜ 0 , P˜ 1  = 0. We then construct an orthogonal set of polynomials {P˜ 0 , P˜ 1 , . . . , P˜ k } by mathematical induction. Suppose {P˜ 0 , P˜ 1 , . . . , P˜ k } form an orthogonal set with P˜ m ∈ m , P˜ m = 0, m = 0,

9.7 Approximation Theory

521

1, 2, . . . , k. We determine ak and bk so that P˜ k+1 (x) = (x − ak )P˜ k (x) − bk P˜ k−1 (x)

(9.135)

is orthogonal to P˜ 1 , P˜ 2 , . . . , P˜ 2 . Then (9.135) implies that % $ % $ % $ % $ P˜ k+1 , P˜ m = xP˜ k , P˜ m − ak P˜ k , Pm − bk P˜ k−1 , P˜ m . If m < k − 1, the inductive hypothesis implies that P˜ k−1 , P˜ m  = P˜ k , P˜ m  = 0. Furthermore, xP˜ k , P˜ m  = P˜ k , xP˜ m . In view of the fact that xP˜ m ∈ k−1 , we can write xP˜ m = C0 P˜ 0 + C1 P˜ 1 + · · · + Ck−1 P˜ k−1 , and this leads to P˜ k , xP˜ m  = 0 and hence P˜ k+1 , Pˆ m  = 0, m = 0, 1, . . . , k − 2. Also, % $ % $ % $ P˜ k+1 , P˜ k = xP˜ k , P˜ k − ak P˜ k , P˜ k . If we choose ak =

xP˜ k , P˜ k  , P˜ k , P˜ k 

(9.136)

then P˜ k+1 , P˜ k  = 0. On the other hand, % $ % $ % $ P˜ k+1 , P˜ k−1 = xP˜ k , Pk−1 − bk P˜ k−1 , P˜ k−1 . If we choose bk =

xP˜ k , P˜ k−1  , P˜ k−1 , P˜ k−1 

(9.137)

then P˜ k+1 , P˜ k−1  = 0. Since P˜ k+1 ∈ k+1 , {P˜ 0 , P˜ 1 , . . . , P˜ k+1 } is an orthogonal set. It follows from the mathematical induction that {P˜ 0 , P˜ 1 , . . . , P˜ n }, P˜ m ∈ m , is a set of orthogonal polynomials obtained from P˜ 0 = 1 and

P˜ 1 = (x − a0 )

by the three-term recurrence relation (9.135) for k = 0, 1, . . . , n − 1, where ak is given by (9.136) for k = 0, 1, 2, . . . , n − 1, and bk by (9.137) for k = 1, 2, 3, . . . , n − 1. We next introduce Pk =

P˜ k , P˜ k

k = 0, 1, . . . , n,

(9.138)

so that {P0 , P1 , . . . , Pn } forms an orthonormal set with respect to the weight function w(x). Also, Pk ∈ k and the leading coefficient of Pk is positive for

522

Chapter 9

Miscellaneous Applications

k = 0, 1, . . . , n. Thus, the set is unique. Finally, the least-square polynomial approximation of degree n of f is then determined by (9.131), where Pk is given by (9.138) and λk =  f , Pk ,

k = 0, 1, . . . , n.

(9.139)

To find Q∗n+1 (x), we need only find Pn+1 (x) by (9.135) and (9.138) and then calculate λn+1 by (9.139). The final result is Q∗n+1 = Q∗n + λn+1 Pn+1 .

(9.140)

As a concluding remark, we must mention that a lot of progress has been made on the theory of least-squares approximations by using various orthogonal polynomials in −1 ≤ x ≤ 1. These include Jacobi polynomials Pn(α,β) (x), (w(x) = (1 − x)α (1 + x)β , α, β > −1); Legendre polynomials Pn (x) (α = β = 0, w(x) = 1); and Chebyshev polynomials Tn (x) (α = β = − 12 , w(x) = (1 − x2 )−1/2 ). Other orthogonal polynomials are also of interest and can be obtained from the Chebyshev polynomials Tn (x), which satisfy the recurrence relation Tn (x) = 2xTn−1 (x) − Tn−2 (x),

n ≥ 2,

with T0 (x) = 1 and T1 (x) = x. It follows from Tn (x) = cos nθ (n = 0, 1, 2, . . .), where x = cos θ , 0 ≤ θ ≤ π , that Tn (x) = n sin nθ/ sin θ . We then define the new polynomials Un (x) of degree at most n by Un (x) =

sin(n + 1)θ , sin θ

n = 0, 1, 2, . . . ,

(9.141)

where x = cos θ . These are called the Chebyshev polynomials of the second kind. It is easy to check that polynomials (9.141) are orthogonal with respect to w(x) = (1 − x2 )1/2 and hence are constant multiples of Jacobi’s polynomials Pn(1/2,1/2) (x). Using L’Hôpital’s rule, it follows that Un (1) = (n + 1), and then 1 · 3 · 5 · · · (2n + 1) Un (1). 2n (n + 1)! There are many identities connecting Tn (x) and Un (x). Some of them are given as exercises. Pn(1/2,1/2) (1) =

9.8 The Shannon Sampling Theorem An analog signal f (t) is a continuous function of time t defined in −∞ < t < ∞ with the exception of perhaps a countable number of jump discontinu-

9.8 The Shannon Sampling Theorem

523

ities. Almost all analog signals f (t) of interest in engineering have finite energy. By this we mean that f ∈ L2 (−∞, ∞). The norm of f , , f =

∞ −∞

 f (t)2 dt

-1/2 ,

represents the square root of the total energy content of the signal f (t). The spectrum of a signal f (t) is represented by its Fourier transform F(ω), defined by (5.106), where ω is called the frequency. The frequency is measured by ν = ω 2π in terms of Hz. A signal f (t) is called band-limited if its Fourier transform F(ω) has a compact support, that is F(ω) = 0 for |ω| > ω0

(9.142)

for some ω0 > 0. If ω0 is the smallest value for which (9.142) holds, then it is called the bandwidth of the signal. Even if an analog signal f (t) is not bandlimited, we can reduce it to a band-limited signal by what is called an ideal lowpass filtering. To reduce f (t) to a band-limited signal fω0 (t) with bandwidth less than or equal to ω0 , we consider  F(ω) for |ω| ≤ ω0 , (9.143) Fω0 (ω) = 0 for |ω| > ω0 , and find fω0 (t), called the low-pass filter function, by the inverse Fourier transform  ∞  ω0 1 1 fω0 (t) = eiωt Fω0 (ω)dω = eiωt Fω0 (ω)dω. 2π −∞ 2π −ω0 In particular, if

 Fω0 (ω) =

1 0

for |ω| ≤ ω0 , for |ω| > ω0 ,

(9.144)

then Fω0 (ω) is called the gate function and its associated signal fω0 (t) is given by  ω0 1 sin ω0 t . (9.145) fω0 (t) = eiωt dω = 2π −ω0 πt This function is called the Shannon sampling function. When ω0 = π , fπ (t) is called the Shannon scaling function. Both fω0 (t) and Fω0 (ω) are shown in Figure 9.1. In engineering, linear analog filtering is defined by the time-domain convolution. If ϕ(t) is the filter function, then the input-output relation of this filter is given by  ∞ g(t) = (ϕ ∗ f )(t) = ϕ(τ )f (t − τ )dτ. (9.146) −∞

524

Chapter 9

Miscellaneous Applications

Figure 9.1 The low-pass filter function and the low-pass filter characteristic. In the frequency domain, the filtering process is represented by pointwise multiplication as G(ω) = (ω)F(ω),

(9.147)

where G(ω) is the Fourier transform of g(t) and (ω) is the Fourier transform of ϕ(t), usually called the transfer function of the filter. Consider the limit, as ω0 → ∞, of the Fourier integral  ∞ e−iωt fω0 (t)dt 1 = lim Fω0 (ω) = lim ω0 →∞

= lim

ω0 →∞ −∞

∞

ω0 →∞ −∞ ∞ −iωt

 =



−∞

e

e−iωt

sin ω0 t dt = πt



∞

−∞

sin ω0 t dt ω0 →∞ πt

e−iωt lim

(9.148)

δ(t)dt.

Clearly, the delta function can be thought of as the limit of the sequence of signal functions fω0 (t). More explicitly, sin ω0 t . ω0 →∞ πt

δ(t) = lim

The band-limited signal fω0 (t) has the representation  ω  ∞ 1 1 iωt F(ω)e dω = F(ω)Fω0 (ω)eiωt dω, fω0 (t) = 2π −ω0 2π −∞

(9.149)

9.8 The Shannon Sampling Theorem

525

which gives, by the Convolution Theorem 5.11.11,  fω0 (t) =

∞ −∞

 f (τ )fω0 (t − τ )dτ =

∞

sin ω0 (t − τ ) f (τ )dτ. π(t − τ ) −∞

(9.150)

This gives the sampling integral representation of a band-limited signal fω0 (t). Next, we consider the Fourier series expansion of the Fourier transform Fω0 (ω) of a band-limited signal fω0 (t) on the interval −ω0 < ω < ω0 in terms ω of the orthogonal set of functions (exp(− inπ ω0 )) in the form   inπ ω an exp − , Fω0 (ω) = ω0 n=−∞ ∞ 

(9.151)

where the Fourier coefficients an are given by an =

1 2π



    inπ ω 1 nπ f ω0 Fω0 (ω) exp − dω = . ω0 2π ω0 −ω0 ω0

Thus, the Fourier series expansion (9.151) reduces to the form     ∞ nπ 1  inπ ω Fω0 (ω) = fω exp − . 2ω0 n=−∞ 0 ω0 ω0

(9.152)

Multiplying (9.152) by eiωt and integrating over (−ω0 , ω0 ) leads to the reconstruction of the signal function fω0 (t) in the form  fω0 (t) =

ω0 −ω0

1 = 2π

Fω0 (ω)eiωt dω



ω0

−ω0

&

∞ 

n=−∞

 f ω0

nπ ω0



 ' inπ ω exp − dω ω0

,     ω0 ∞ nπ nπ 1  fω exp iω t − dω = 2π n=−∞ 0 ω0 ω0 −ω0 =

∞  n=−∞

 f ω0

nπ ω0



sin ω0 (t − ω0 (t −

nπ ω0 ) . nπ ω0 )

(9.153)

This formula is referred to as the Shannon sampling theorem. It represents  an  expansion of a band-limited signal fω0 (t) in terms of its discrete values fω0 nπ ω0 . This is very important in practice, because most systems receive discrete samples as an input.

526

Chapter 9

Miscellaneous Applications

9.9 Linear and Nonlinear Stability We consider linear and nonlinear problems of stability and instability for differential systems. In dynamical systems, the state at any time t can be represented by an element of a Banach (or Hilbert) space E. Suppose that the dynamics of a physical system are governed by the evolution equation du = F(λ, u, t), dt

(9.154)

where λ ∈  is a parameter,  is a set of parameters (for instance  = R), u is a function of a real variable t with values in E, and F is a mapping from  × E × R into E.

Definition 9.9.1.

(Autonomous dynamical system) The dynamical system governed by (9.154) is called autonomous if the function F does not depend explicitly on t . For autonomous systems, (9.154) can be written in the form du/dt = F(λ, u).

Definition 9.9.2.

(Equilibrium solution) If F(λ0 , u0 ) = 0 for some λ = λ0 and u = u0 , then u0 is called an equilibrium solution.

Definition 9.9.3.

(Stable, unstable, and asymptotically stable solutions) Let u0 be an equilibrium solution of equation (9.154). (a) u0 is called stable if for every ε > 0 there exists a δ > 0 such that   u(t) − u0  < ε for all solutions u(t) of (9.154) such that u(0) − u0 < δ . (b) u0 is called unstable if it is not stable. (c) u0 is called asymptotically stable if it is stable and u(t) − u0 → 0 as t → ∞.

Example 9.9.4. Consider the scalar equation x˙ = 0. Every solution of this equation has the form x = c, where c is a constant. Thus, every solution is stable but not asymptotically stable.


9.9 Linear and Nonlinear Stability

527

Example 9.9.5. Consider the system du/dt = λu, u(0) = u0 , where u(t) is real

for each t and λ ∈ R. This equation has the equilibrium solution u0 (t) = 0. The general solution is u(t) = u0 eλt .

If λ ≤ 0, then the zero solution is stable. If λ > 0, the solution is unstable because u(t) → ∞, no matter how small u0 is.


Example 9.9.6. Consider the equation x˙ = x2 with x(0) = x0 . The solution of this equation is obtained by separating the variables and has the form x(t) =

x0 . 1 − x0 t

The solution is not defined for t = 1/x0 . Thus, x(t) ≡ 0 is a solution which is unstable.


Example 9.9.7. Consider a linear autonomous system u˙ = Lu + v,

(9.155)

where u(t) ∈ E for each t, L : E → E is a linear operator which does not depend on t, and v is a given element of E. Clearly, u0 ∈ E is an equilibrium solution of (9.155) if Lu0 = −v. We suppose the solution of (9.155) is of the form u(t) = u0 + eλt w, where λ is a constant and w ∈ E. Clearly, u(t) satisfies (9.155) provided Leλt w = λeλt w. This means that λ is an eigenvalue of L with eigenvector w. If the eigenvalue λ has a positive real part and w is a normalized eigenvector, then, for any ε > 0, the function u(t) = u0 + εweλt is a solution of (9.155) such that u(0) − u0 = ε and u(t) − u0 → ∞ as t → ∞. This shows that the equilibrium solution u0 is unstable provided there is an eigenvalue with a positive real part.
This example leads to the Principle of Linearized Stability, which can be described as follows: Consider a system of ordinary differential equation u˙ = F(λ, u),

(9.156)

where u = (u1 , u2 , . . . , un ), F = (F1 , F2 , . . . , Fn ), and λ is a parameter. Let u0 be the equilibrium solution with λ = λ0 , so that F(u0 , λ0 ) = 0. Suppose the solution of (9.156) can be written as u(t) = v(t) + u0 , where v(t) is the perturbation from equilibrium. It follows from u˙ = F(u, λ0 ) that ,   ∂Fi (v) + O v 2 v˙ = u˙ = F(v + u0 , λ0 ) = F(u0 , λ0 ) + ∂uj

528 or

Chapter 9

Miscellaneous Applications

  v˙ = Av + O v 2 ,

(9.157)

where A = [∂Fi /∂uj ](u0 ,λ0 ) and G(v) = O( v ) represent a term such that   G(v) ≤ c v 2 , 2

where c is a constant. Neglecting the second term in (9.157), we obtain the linear equation v˙ = Av.

(9.158)

v(t) = etA u0 .

(9.159)

The solution of this equation is

Clearly, all solutions of this equation decay if the spectrum of A lies in the left half-plane. Some solutions of (9.159) may grow exponentially provided A has eigenvalues in the right half-plane. In general, the second order term is negligible when the perturbations are small. This heuristic argument can be justified by Lyapunov’s theorem (Aleksander Mikhailovich Lyapunov (1857– 1918)):

Theorem 9.9.8. (Lyapunov’s theorem) If all eigenvalues of A have negative real parts, then u0 is a stable equilibrium solution of (9.156). If some eigenvalues of A have positive real parts, then u0 is an unstable solution. A rigorous proof of this theorem is beyond the scope of this book. However, the reader is referred to Coddington and Levinson (1955). The following example shows that the weak inequality Re(λ) ≤ 0 for all eigenvalues does not ensure stability.

Example 9.9.9. Consider the equation u˙ = Au, where u(t) ∈ R2 and A is the  

matrix operator 00 01 . If u0 is an equilibrium solution of this equation, then Au0 = 0. Clearly, u0 = (a, 0) represents an equilibrium solution for any number a. The only eigenvalue of A is zero. If we write u = (x, y), then the given equation becomes x˙ = y and y˙ = 0. Hence, the general solution is y = m, x = mt + c, where m and c are constants. For sufficiently small m and c, the solution u(t) = (mt + c, m) can be made sufficiently close to u0 = (a, 0) at t = 0. But u(t) − u0 → ∞ as t → ∞. This shows that the equilibrium solution is unstable.


Theorem 9.9.10. (Stability criterion) If A is a linear operator on a space E and

A + A∗ is negative semidefinite, that is v, (A + A∗ )v ≤ 0 for all v ∈ E, then all equilibrium solutions of the equation u˙ = Au + f

(9.160)

9.9 Linear and Nonlinear Stability

529

are stable, where u is an element of a Hilbert space E for each t, A : E → E is independent of t, and f is a given element of E.

Proof: Suppose u0 is an equilibrium solution of (9.160), that is Au0 = 0, and u(t) is any other solution. If v = u − u0 , then v˙ = Av. Thus, d2 d2 2 v = v, v = v, v ˙ + v, ˙ v dt 2 dt 2 = v, Av + Av, v $   % = v, A + A∗ v . If A + A∗ is negative semidefinite, then d2 v 2 ≤ 0. dt 2 This means that v is a non-increasing function. Consequently, if u(0) − u0 < ε, then u(t) − u0 < ε for all t > 0. This shows that all equilibrium solutions are stable.
We next consider the stability of a general nonlinear autonomous equation u˙ = Nu.

(9.161)

The question of stability of an equilibrium solution u0 of (9.161) is concerned with the effects of small initial displacements of u from u0 , and it only involves values of u in the neighborhood of u0 . If N is Fréchet differentiable, then the operator N can be approximated by the linear operator N (u) in the neighborhood of u0 , and linear stability theory can be used. Hence, Nu = Nu0 + N (u0 )(u − u0 ) + o(u − u0 ),

(9.162)

where Nu0 = 0. Neglecting the term o(u − u0 ), equation (9.162) is approximately equal to u˙ = N (u0 )(u − u0 ).

(9.163)

This equation may be called the linearized approximation of the nonlinear equation (9.161). Its stability can be determined by stability criteria discussed earlier. When u is near u0 , (9.163) is the linearized approximation to (9.161), so it is naturally assumed that the stability of the linearized equations determines that for the nonlinear equations. This principle is generally accepted as valid in the applied literature, and stability is determined formally by solving the associated linear eigenvalue problem. However, this general principle is not necessarily true as shown by a counterexample.

530

Chapter 9

Miscellaneous Applications

Example 9.9.11. Consider the nonlinear equation u˙ = u3 , where u(t) ∈ R for

each t. The equilibrium solution is u0 = 0. It can be explicitly solved by using the initial condition u(0) = u0 , and the solution is u2 =

u20 , 1 − 2u20 t

which is not defined for t = 2u1 2 . Thus, u0 is an unstable equilibrium. However, 0 the linearized equation u˙ = 0 admits a stable solution. Thus, the stability of the linearized equation does not imply stability of the nonlinear equation. The difficulty associated with this example is that the linearized equation has eigenvalue λ = 0 (critical case when Re λ = 0). In other words, the linearized system is only marginally stable. This means that an arbitrarily small perturbation can push the eigenvalue into the right half-plane, and make the system unstable. The eigenvalue zero corresponds to a constant solution of the linearized equation, and an arbitrarily small perturbation can change this constant solution and thus lead to instability. However, if all the eigenvalues of a linearized problem are negative, then its solutions tend to u0 exponentially. The small perturbations involved in going from the linearized to the nonlinear problem cannot change exponential decay of u − u0 into growth, so in this case the nonlinear problem will be stable.


9.10 Bifurcation Theory Bifurcation is a phenomenon involved in nonlinear problems and is closely associated with the loss of stability. Section 9.9 showed that the stability of a dynamical system depends on whether the eigenvalues of the linearized operator are positive or negative. These eigenvalues correspond to bifurcation points. We shall discuss bifurcation theory in terms of operator equations in a real Banach (or Hilbert) space. By a nonlinear eigenvalue problem, we usually mean the problem of determining appropriate solutions of a nonlinear equation of the form F(λ, u) = 0,

(9.164)

where F : R × E → B is a nonlinear operator, depending on the parameter λ, which operates on the unknown function or vector u, and E and B are real Banach (or Hilbert) spaces. Bifurcation theory deals with the existence and behavior of solutions u(λ) of equation (9.164) as a function of the parameter λ. Of particular interest is the process of bifurcation (or branching) where a given solution of (9.164) splits into two or more solutions as λ passes through a critical value λ0 , called a bifurcation point.

9.10 Bifurcation Theory

531

Figure 9.2 Bifurcation diagram.

Definition 9.10.1.

(Bifurcation point) The solution of (9.164) is said to bifurcate from the solution u0 (λ0 ) at the value λ = λ0 if the equation has at least two distinct solutions u1 (λ) and u2 (λ) such that they tend to u0 ≡ u0 (λ0 ) as λ → λ0 . The points (λ0 , u0 ) satisfying equation (9.164) are referred to as bifurcation (or branch) points if, in every neighborhood of (λ0 , u0 ), there exists a solution (λ, u) different from (λ0 , u0 ). The first problem of bifurcation theory is to determine the solution u0 and the parameter λ0 at which bifurcation occurs. The second problem is to find the number of solutions which bifurcate from u0 (λ0 ). The third problem is to study the behavior of these solutions for λ near λ0 . To illustrate bifurcation, we consider the linear eigenvalue problem Lu = λu,

(9.165)

where L is a linear operator acting on a function or a vector u in some Banach space and λ ∈ R. For every value of λ, (9.165) has a trivial solution u = 0 with the norm u = 0. Suppose there is a sequence of eigenvalues λ1 < λ2 < λ3 < · · ·, and the corresponding normalized eigenfunctions u1 , u2 , u3 , . . . such that Luk = λk uk ,

uk = 1,

k = 1, 2, 3, . . . .

(9.166)

Then, for any real number a, nontrivial solutions are u = auk , k = 1, 2, 3, . . . , with the norm u = a. The norms of both trivial and nontrivial solutions are shown graphically by Figure 9.2. Many examples of bifurcation phenomena occur in both differential and integral equations. One such example is as follows:

Example 9.10.2. Consider a thin elastic rod with pinned ends lying in the x– z plane. The shape of the rod is described by two functions u(x) and w(x), which are the dimensionless displacement functions in the x and z directions. The x-displacements of its end points are prescribed. The displacement func-

532

Chapter 9

Miscellaneous Applications

tions u(x) and w(x) satisfy the following differential equations and boundary conditions: d2 w + λw(x) = 0, 0 ≤ x ≤ 1, dx2     1 dw 2 du + = −µλ, 0 ≤ x ≤ 1, dx 2 dx w(0) = w(1) = 0,

u(0) = −u(1) = a > 0,

(9.167) (9.168) (9.169)

where the parameter λ is proportional to the axial stress in the rod, the constant a in (9.169) is proportional to the prescribed end displacement and is referred to as the end-shortening, and µ is a positive physical constant. Consider the linearized problem where the nonlinear term wx2 is absent. The solution of the linearized equation (9.168) is u(x) = a(1 − 2x),

(9.170)

where a = λµ/2. The solution of (9.167) and (9.169) is w(x) ≡ 0 unless λ is an eigenvalue λn given by λ = λn = n2 π 2 ,

n = 1, 2, 3, . . . .

(9.171)

In this case, w is a multiple of the eigenfunctions wn given by w(x) = An wn (x) = An sin nπ x,

n = 1, 2, 3, . . . ,

(9.172)

where the An are constants. From a = 12 λµ and λ = λn = n2 π 2 , we conclude that if a = an = 12 µλn , then the rod buckles into a shape given by (9.170) and (9.172) with an undetermined amplitude An . The numbers an are called the critical end-shortenings. For a = an , n = 1, 2, . . . , the rod remains straight because the solution of (9.167) and (9.169) is w(x) ≡ 0.

(9.173)

We now consider the nonlinear problem (9.167) to (9.169). The solution of the problem is still given by (9.172) when λ = λn and by (9.173) when λ = λn . To find u(x) when λ = λn , we put (9.172) into (9.168) and integrate using (9.169) at x = 0 to obtain   1 A2n x + nπ A2n sin 2nπ x. u(x) = un (x) = a − µλn 1 + (9.174) 4µ 8 In view of the boundary condition u(1) = −a, we obtain   A2n . a = an 1 + 4µ

(9.175)

9.10 Bifurcation Theory

533

Figure 9.3 Bifurcation diagram for the thin rod.

This is a relation between the end-shortening and the amplitude. The bifurcation diagrams for the thin rod are given in Figure 9.3. The diagram shows that, for a < a1 , the only solution is the trivial solution w ≡ 0. At a = a1 , the nontrivial solution w1 = A1 sin π x bifurcates from the trivial solution and continues to exist for all a > a1 . The point a = a1 is called the first bifurcation point, and the nontrivial solution is called the first bifurcation solution. For each n, nontrivial solutions of (9.175) for An are possible if and only if a ≥ an . The solutions bifurcate from the trivial (unbuckled) state An = 0 at a = an . Thus, the solution of the linearized problem determines the bifurcation points of the nonlinear problem. For any a in an ≤ a ≤ an+1 , there are 2n + 1 solutions. For a < a1 , no buckling is possible. We also note from (9.175) that da/dAn = an An /2µ. Hence, for a fixed amplitude A, the parabola in Figure 9.3 bifurcating from an has a steeper slope than that bifurcating from am if m < n. Clearly, these parabolas do not intersect. For any fixed value of a, the bifurcation solutions can be classified by the values of the potential energy associated with them. We also observe that the potential energy is equal to the internal energy, since the displacements are specified at the ends of the rod. Consequently, the potential energy is proportional to the functional V defined by 1 V(w) = 2

 1& 0

2 wxx

 ' 1 1 2 2 ux + wx dx. + µ 2

(9.176)

In the unbuckled state, equations (9.170) and (9.173) hold with a = λµ/2, and the corresponding potential energy is  2 2a 1 . V∞ = µ 2 µ

(9.177)

534

Chapter 9

Miscellaneous Applications

The potential energy Vn of the buckled state given by (9.172) is obtained by substituting (9.171), (9.172), and (9.174) into (9.176) in the form Vn = λn (2a − an ).

(9.178)

Hence, 2 Vn − V∞ = − (a − an )2 ≤ 0, a ≥ an , µ   2 (Vn − Vm ) = (an − am ) (a − an ) + (a − am ) ≥ 0, µ

(9.179) a ≥ an ≥ am . (9.180)

It follows from (9.179) and (9.180) that, for fixed a > a1 , the straight state has the largest energy, and the branch originating from a1 has the smallest energy. For fixed a in the interval an ≤ a ≤ an+1 , the energies of the branches are ordered as V∞ > Vn > Vn−1 > · · · > V1 . For the state of smallest energy, the displacement function of this state is   a √ w = A1 w1 = ±2 µ − 1 sin π x for all a > an . (9.181) a1


Suppose the solutions of (9.164) represent equilibrium solutions for a dynamical system which evolves according to the time-dependent equations ut = F(λ, u),

(9.182)

where u : R → E and E is a Banach (or Hilbert) space. An equilibrium solution u0 is stable if small perturbations from it remain close to u0 as t → ∞; u0 is asymptotically stable if small perturbations tend to zero as t → ∞ (see Section 9.9). When the parameter λ changes, one solution may persist but become unstable as λ passes a critical value λ0 , and it is at such a transition point that new solutions may bifurcate from the known solution. One of the simple nonlinear partial differential equations which exhibits the transition phenomena shown in Figure 9.3 is ut = ∇ 2 u + λu + u3

in D,

(9.183)

u=0

on ∂D,

(9.184)

where D is a smooth bounded domain in RN . The equilibrium states of (9.183) are given by solutions of the time-independent equation (ut ≡ 0). One solution is obviously u = 0, which is valid for all λ; this solution becomes unstable at λ = λ1 , the first eigenvalue of the Laplacian: ∇ 2 u1 + λ1 u1 = 0 on D, u1 = 0 on ∂D. For λ > λ1 , there are at least three solutions of the nonlinear equilibrium equation. The nature of the solution set in the neighborhood of (λ1 , 0)

9.11 Exercises

535

Figure 9.4 Bifurcation diagram where unstable solutions are represented by dashed lines. is given in Figure 9.4; the new bifurcating solutions are stable. The Laplacian has a set of eigenvalues λ1 < λ2 < λ3 < · · · , which tend to infinity, and all of these eigenvalues are potential bifurcation points. In the theory of calculus in Banach spaces, the following version of the Implicit Function Theorem is concerned with the existence, uniqueness, and smoothness properties of the solution of the equation (9.164).

Theorem 9.10.3. (Implicit Function Theorem) Suppose , E, B are real Banach spaces and F is a Fréchet differentiable mapping from a domain D ⊂  × E to B. Assume F(λ0 , u0 ) = 0 and the Fréchet differential F (λ0 , u0 ) is an isomorphism from E to B. Then locally, for λ − λ0 sufficiently small, there is a differentiable mapping u(λ) from  to E, with (λ, u(λ)) ∈ D, such that F(λ, u(λ)) = 0. Moreover, (λ, u(λ)) is the only solution of F = 0 in a sufficiently small neighborhood D ⊂ D. If F is Cn then u is Cn . If λ, E, and B are complex Banach spaces and F is Fréchet differentiable, then F is analytic and u is analytic in λ. The proof of the theorem is beyond the scope of this book. However, the theorem can be proved by using a contraction mapping argument and is adequate for most physical applications. The reader is referred to Sattinger (1973) or Dieudonné (1969) for a detailed discussion of proofs. Bifurcation phenomena typically accompany the transition to instability when a characteristic parameter crosses a critical value, and hence they play an important role in applications to mechanics. Indeed, the area of mechanics is a rich source of bifurcation and instability phenomena, and the subject has always stimulated the rapid development of functional analysis. For more on bifurcation theory, see Arrowsmith and Place (1992) and Guckenheimer and Holmes (1983).

9.11 Exercises 1. Show that the norm in an inner product space is strictly convex.

536

Chapter 9

Miscellaneous Applications

2. Let H1 and H2 be real Hilbert spaces. Show that if T is a bounded linear operator from H1 into H2 , and f is a real functional on H1 defined by f (x) = u − Tx 2 , where u is a fixed vector in H2 , then f has a Fréchet derivative at every point given by f (x) = −2T ∗ u + 2T ∗ Tx, where T ∗ is the adjoint of T. 3. Suppose T : B1 → B2 is Fréchet differentiable on an open set  ⊂ B1 . Show that if x ∈  and h ∈ B1 are such that x +th ∈  for every t ∈ [0, 1], then     T(x + h) − T(x) ≤ h sup T (x + αh). 0 0, λ = 0 and λ < 0 for the second-order nonlinear systems: (a) x˙ = λ − x2 , y˙ = y. (b) x˙ = λx − x2 , y˙ = −y. (c) x˙ = λx ± x3 , y˙ = −y. Show that system (a) represents a saddle-node bifurcation and that systems (b) and (c) describe supercritical and subcritical bifurcations. 61. The Lorenz equations (1963) represent a system of three autonomous ordinary differential equations 8 x˙ = − x + yz, 3

y˙ = −10(y − z),

z˙ = −xy + 28y − z.

9.11 Exercises

545

This system describes as the leading order approximation to the behavior of an ideal model of the Earth’s atmosphere. Find the equilibrium points of the system.

This page intentionally left blank

Hints and Answers to Selected Exercises

1.7 Exercises 8. Consider the function f (x) = 1p x + 1q − x1/p . Show that f is decreasing on [0, 1] and that f (1) = 0. 9. If |x| < 1, then |x|q ≤ |x|p . Consider the sequence (n−1/p ). 10. The system λ1 = x 1 λ1 + λ2 = x 2 λ1 + λ2 + λ3 = x 3 has a solution for any (x1 , x2 , x3 ) ∈ R3 . 12. Prove that if λ0 + λ1 x + · · · + λn xn = 0 for all x, then λ0 = λ1 = · · · = λn = 0. 14. See Exercise 12 or 13. 15. (1, 0, 0, . . .), (0, 1, 0, 0, . . .), (0, 0, 1, 0, 0, . . .), . . . . 16. If x = α1 e1 + · · · + αn en and x = β1 e1 + · · · + βn en , then (α1 − β1 )e1 + · · · + (αn − βn )en = 0. 17. d1 + d2 + · · · + dn . 18. Use the triangle inequality a + b ≤ a + b with a = x − y and b = y.

547

548

Hints and Answers to Selected Exercises

26. No. If fn (x) = max{0, 1 − |x − n|}, then αn fn → 0 as n → ∞ for any sequence of scalars αn . 28. Consider xn = (1, 12 , 13 , . . . , n1 , 0, 0, . . .). If 1 ≤ p < q, the convergence in lp implies convergence in lq , but not conversely. 29.   N N    

|z |2 ≤ |z | ≤ N max |z |, . . . , |z | . max |z |, . . . , |z | ≤  1

N

n

n

n=1

1

N

n=1

30.



  f 1 = max f (x),

1

 f (x)dx.

f 2 =

x∈[0,1]

0

38. No. 41. Modify the proof in Example 1.4.6. 42. Consider the space of polynomials on [0, 1] with the norm   f = max f (x). x∈[0,1]

43. Every sequence (xn ) convergent to 0 has a subsequence (xpn ) such that xpn ≤ 1/2n . 44. No. 45.

 x       max  f (t)dt  ≤ max f (x). x∈[0,1] x∈[0,1] 0

46. See Exercise 52. 52. No. 53. (0, 0), (1, 1), (−1, −1). 54. No. 55. max

1≤k≤N

N 

|akn | < 1.

n=1

2.16 Exercises   3. Show that if λ1 f1 + · · · + λn fn = 0, then λ1 f1 + · · · + λn fn = 0. 5. If gk denotes the characteristic function of [ak , bk ), then −M(g1 + · · · + gn ) ≤ f ≤ M(g1 + · · · + gn ).

2.16 Exercises

549

8. The proof is easy if f is the characteristic function of an interval [a, b). Then generalize for step functions. 10. Expand f into a series of step functions, and then replace every step function in that series by its basic representation. 13. No. 16. First prove it for a positive function. Then use the representation f = f + − f − , where f + and f − are as defined in Exercise 15. 17. Expand the characteristic function of an interval [a, b) into a series of tent functions, and conversely. 22. See Exercise 17. 24. f (x) =

∞ 

ne−n (x−n) . 6

2

n=1

27. If f  f1 + f2 + · · · and f1 , f2 , . . . are step functions, then f1 g, f2 g, . . . are integrable and fg  f1 g + f2 g + · · · . 29. Use the Lebesgue dominated convergence theorem. 31. Redefine fn ’s at those points x where |fn (x)| > h(x). 36. Show that for

 fn (x) =

f (x) if x ∈ An , 0 if x ∈ / An ,

we have fn → 0 a.e. Then use the Lebesgue dominated convergence theorem. 46. Note that (f ∗ g)(x + h) − (f ∗ g)(x) = h

 f (t)

g(x + h − t) − g(x − t) dt, h

and then prove that  lim

h→0

=

f (t) 

g(x + h − t) − g(x − t) dt h

f (t) lim

47. Compare with Exercise 46.

h→0

g(x + h − t) − g(x − t) dt. h

550

Hints and Answers to Selected Exercises

3.8 Exercises 4. (a) No. (b) It is an inner product space, but it is not a Hilbert space. 5. It is an inner product space, but it is not a Hilbert space. 7. Prove, using Schwarz’s inequality, that x + y = x + y if and only if x and y are linearly independent. 9. Use x ± αy, x ± αy = x 2 ± αx, y ± αy, x + |α|2 y 2 . 11. Use Schwarz’s inequality and the fact that x = x, x /x if x = 0. 12. See, for example, Friedman (1982), Theorem 6.1.5. 13. No. The space does not satisfy the parallelogram law. 15. Use the parallelogram law or verify by direct calculations. 18. Yes. Use continuity of the inner product. 19. For (c) and (d), use an orthonormal sequence. √

34. a = c = 0, b = 56 .  36. x = ∞ n=1 αn an . 46. 1 , √ b−a . nπ(2x − b − a) 2 cos , b−a b−a . nπ(2x − b − a) 2 sin , b−a b−a

n = 1, 2, . . . , n = 1, 2, . . . .

47. The set of all functions in L2 (R) such that f (−x) = −f (x) almost everywhere. 48. The set of all functions in L2 ([−π, π ]) such that f (−x) = f (x) almost everywhere. 49. Use Theorem 3.6.2. 52. Consider, for example, the set S = {(x, 0): x ∈ R} and the point x0 = (0, 1).  53. y = ∞ n=1 x, en en .

4.12 Exercises

551

55. Consider, for example, xn (t) = ne−n (t−t0 ) . 6

2

58. The sequence (1, 0, 0, . . .), (0, 1, 0, 0, . . .), (0, 0, 1, 0, 0, . . .), . . . is a complete orthonormal system in l2 .

4.12 Exercises 2. Define ϕ(x, y) = Ax, y and use the polarization identity. 5. Compare with the proof of Schwarz’s inequality. 7. λ1 = 4, x1 = (2, 3), λ2 = −1, x2 = (1, −1). 8. T ∗ (x, y) = (x + 2y, 3x + y). 28. No. Consider, for example, T : R2 → R2 defined by √   3x x y y T(x, y) = + , − √ . 2 3 2 3 3 32. See, for example, Friedman (1982), Lemma 6.6.6. √ 34. I = I . 37. Use Theorem 4.6.14. 39. No. Consider, for example, P = −I . 41. Find an example in R2 . 50. Use Corollaries 4.8.13 and 4.10.2. 55. See Example 4.5.18.

5.13 Exercises 1

e−sx cos(αf (s))ds is a contraction in C ([0, 1]). √ √ √ √ √ 7. Set c2 = 1. Then c1 = 3, λ1 = 23 , f1 (x) = 23 ( 3x + 1); c1 = − 3, √ √ √ λ2 = 23 , f1 (x) = 23 ( 3x − 1). 5. Show that (Tf )(x) =

0

8. λ1 = π2 , f1 (x) =

,   πc 1 2 2 π π x − 1+ πx , 2 2 4

where c is an arbitrary constant. Similarly, f2 can be found for λ2 = − π2 .

552

Hints and Answers to Selected Exercises

1 12. Apply the method of separable kernel. D(λ) = 12 (12 − 12λ − λ2 ). For D(λ) = 0,     b2 + λ b31 − b22 b1 − λ b21 − b2 , c2 = . c1 = D(λ) D(λ)

The resolvent kernel is (x, t, λ) =

D(x, t, λ) , D(λ)

where &

' 0 M1 (x) M2 (x) D(x, t, λ) = N1 (t) 1 − λa11 −λa12 N2 (t) −λa21 1 − λa22   λ λ . = − (x + t) − (x + t) + λtx + 2 3 Answer: f (x) = ϕ(x) + λ

1 0

(x, t, λ)dt.

14. (a) λ1 = λ2 = π1 , f (x) = c1 cos x + c2 sin x, √ √ √ √ (b) λ1 = 23 i, f1 (x) = 1 − i 3x, λ2 = − 23 i, f2 (x) = 1 + i 3x. 15. (d) f (x) = −2. 17. (a) f (x) = 12 (3x + 1), (b) f (x) = sin x, (c) f (x) = sinh x, (d) f (x) = e−x . 2

20. 

1

Lu, v = 

0 1

= 

  2 x d x d v(x) e 2 + e u(x)dx dx dx   v(x) ex u (x) dx

0 1

=

uLvdx = u, Lv.

0

22. λn =

(2n − 1)2 , 4

  2n − 1 un (x) = Bn sin x, 2

n = 1, 2, . . . .

23. λn = n2 π 2 , un (x) = Bn sin(nπ ln x), n = 1, 2, . . . . 25. Tu, u = (DpD + q)u, u = DpDu, u + qu, u = −p Du 2 + q u 2 .

6.7 Exercises

553

26. 

b

u, v =

u(x)v(x)r(x)dx. a

29. (a)

(b)

3π 4 ,

(c)

31. (a) π , (b) 36.

2 sin aω π ω ,

2 sin ω 2 π( ω ) .

2π 3 .

   ωE0 E0 R ωE0 −αt −Rt/L + I0 e − α e−αt sin ωt − e cos ωt, I(t) = + AL AL L AL 

where A = (α − R/L)2 + ω2 . 37. Iˆ(ω) =

√C (1 + ikRC 2π

41. (b) F {Tu} =

− LCk2 )−1 .

2 1+k2 F {u}.

Since F is unitary, we have   F {Tu} = Tu ≤ 2 u .

6.7 Exercises 1. If f is a continuous function that is not identically  zero, then there exists a ϕ ∈ C ∞ (RN ) with compact support such that RN f (x)ϕ(x)dx = 0. 5. Use the function

* f (t) = 0−1/t 2 e

if t ≤ 0, if t > 0.

6. (a), (b), (e), (i) Yes. (c), (f), (g), (h), (j) No. (d) No, if (xn ) has a convergent subsequence. 10. f (n) = 2δ (n−1) . 11. Use the Riemann–Lebesgue lemma (Theorem 5.11.6). 13. Note that, for every ε > 0,  lim

n→∞ |x|≥ε

e−n x dx = 0. 2 2

14. Use the Riemann–Lebesgue lemma (Theorem 5.11.6) to show that  ∞ ϕ(x) − ϕ(0) dx = 0. sin nx lim n→∞ −∞ x

554

Hints and Answers to Selected Exercises

15. K =

∞

−∞ ϕ(x)dx,

ϕ1 = ϕ − Kϕ0 .

18. Use the Laplace transform and the double Fourier transforms of u(x, y, t) to obtain    u˜ (k, l, s) = s + K k2 + l2 f˜ (k, l). Application of the inverse transforms combined with the convolution theorem gives the solution. 19. The Green’s function associated with the given equation satisfies the equation Gtt − c2 ∇ 2 G = δ(x)δ(y)δ(z) with the same boundary and initial data. Application of the joint Laplace and the three-dimensional Fourier transform gives   : G = (2π )−3/2 s2 + c2 κ 2 ,

κ 2 = k2 + l2 + m2 .

The joint inverse transforms give  ∞ ∞ ∞ 1 G(x, t) = κ −1 sin(ckt) exp(iκ · x)dκ, (2π )3/2 −∞ −∞ −∞ where κ = (k, l, m). The use of spherical polar coordinates with the polar axis along the vector x, so that κ · x = κr cos θ , r = |x|, θ is the polar angle, and dκ = κ 2 dκ sin θ dθ dφ, gives the Green’s function. 20. Apply the joint Laplace and triple Fourier transform to find : G(k, s) =

1 1 , 3/2 2 2 (2π ) s + c κ 2 + d2

κ 2 = k2 + l2 + m2 .

The joint inverse transform gives the solution  ∞ ∞ ∞ sin αt 1 ei(κ·x) dκ, G(x, t) = (2π )3/2 −∞ −∞ −∞ α where κ = (k, l, m), and α = (c2 κ 2 + d2 )1/2 . In terms of the spherical polar coordinates with the vector x as the polar axis, so that κ · x = κr cos θ , r = |x|, θ is the polar angle, and dκ = κ 2 dκ sin θ dθ dφ, we obtain  ∞  π  2π 1 sin αt 2 dφ dκ exp(iκr cos θ ) sin θ dθ G(x, t) = κ (2π )3 0 α 0 0  ∞ sin αt ∂ 1 (cos κr)dκ =− 2 2π r 0 α ∂r ,   1 ∂ d 22 2 J0 =− c t − r H c2 t 2 − r2 4π rc ∂r c

6.7 Exercises     1 d 22 δ c 2 t 2 − r 2 J0 = c t − r2 2π c c √  d J1 ( dc c2 t 2 − r2 )  2 2 2 − H c t −r . 2c (c2 t 2 − r2 )1/2

555

,

21. Apply the joint Laplace and Fourier transform to obtain   −1  u˜ (k, s) = s2 + c2 k2 sf˜ (k) + g˜ (k) + q˜ (k, s) . The inverse Laplace transform combined with the convolution theorem yields u˜ (k, t) = f˜ (k) cos(ckt) +

1 g˜ (k) sin(ckt) + ck ck



t

sin ck(t − τ )˜q(k, τ )dτ.

0

The use of the inverse Fourier transform leads to the final answer:   1 x+ct 1 g(ξ )dξ u(x, t) = f (x − ct) + f (x + ct) + 2 2c x−ct  x+c(t−τ )  1 t dτ p(ξ, τ )dξ. + 2c 0 x−c(t−τ ) 22. The use of the joint Laplace and Fourier transform gives 1 1 9 G(k, s) = √ . 2 2π s + c2 k2 The inverse Fourier transform yields   1 s G(x, s) = exp − |x| 2cs c and hence the inverse Laplace transform gives the answer: G(x, t) =

   1  1 H(t) H(x + ct) − H(x − ct) = H ct − |x| . 2c 2c

∞ 23. (b) u(x, t) = −∞ G(x, ξ, t)g(ξ )dξ , where G(x, ξ, t) is obtained in part (a). Since G(x, ξ, t) = 2c1 if x − ct < ξ < x + ct and 0 elsewhere,  x+ct we have u(x, t) = 2c1 x−ct g(ξ )dξ . y 29. Define a new function v(x, y) = uy (x, y) so that u(x, y) = v(x, η)dη. Then v(x, y) satisfies the Laplace equation vxx + vyy = 0 with

556

Hints and Answers to Selected Exercises

v(x, 0) = g(x). Then use Example 6.6.3 to find v(x, y) so that  y  ∞   1 u(x, y) = v(x, η)dη = g(ξ ) log (x − ξ )2 + y2 dξ. 2π −∞ 30. u(x, y) =

31.

  ∞ πy f (t)dt 1 π  sin πy 2a a −∞ cosh a (x − t) − cos a   ∞ 1 πy g(t)dt π  + sin πy . 2a a −∞ cosh a (x − t) + cos a

π    ∞ f (t) cosh 2a (x − t) dt πy π  u(x, y) = sin πy . 2a −∞ cosh a (x − t) − cos a

33. 1 ϕ(x, y) = 2a



∞

J0 (ak) ikx−|k|y e dk. k −∞

36. (a) Matching conditions at x = ±a are ψ2 (a) = ψ3 (a) and ψ1 (−a) = ψ2 (−a) that give B sin ak + C cos ak = De−aκ and −B sin ak + C cos ak = Ae−aκ . Matching the derivatives at x = ±a are ψ2 (a) = ψ3 (a) and ψ1 (−a) = ψ2 (−a) which yield Bk cos ak − Ck sin ak = −Dκe−aκ and Bk cos ak + Ck sin ak = Aκe−aκ . Adding and subtracting two sets of matching conditions and putting A + D = A1 and A − D = −A2 give 2B sin ak − A2 e−aκ = 0 and 2Bk cos ak + A2 κe−aκ = 0, 2C cos ak − A1 e−aκ = 0 and 2Ck sin ak − A1 κe−aκ = 0. For nontrivial solutions, we obtain      2 cos ak −e−aκ   2 sin ak −e−aκ       2k cos ak κe−aκ  = 0 and  2k sin ak −κe−aκ  = 0. These give the desired solutions. 37. Use the Fourier transform of gt (x) and the Convolution theorem of the Fourier transform.

7.12 Exercises 1. (a) dtd ∂∂Lx˙ i − ∂x∂ i = 0 implies m¨xi − kxi = 0. Multiply this equation by x˙ i and integrate to obtain 12 m˙xi2 + 12 kxi2 = constant.

557

7.12 Exercises

(b) Use (7.259abc) and L in 1(a) and show that it becomes the expression for L in 1(b). 2. ∂L = mr2 θ˙ , ˙ ∂θ  1 2 where L = T − V = m r˙ + r2 θ˙ 2 − V(r); 2   p2 1 1 p2r + θ 2 + V(r), H = T + V = m r˙2 + r2 θ˙ 2 + V(r) = 2 2 m 2mr pθ ∂H pr ∂H = , = 2. r˙ = θ˙ = ∂pr m ∂pθ mr pr =

∂L = m˙r, ∂ r˙

pθ =

∂H ˙ Then use p˙ r = − ∂H ∂r and pθ = − ∂θ .

4. (a)  2   1 p 1 2 1 p2 2 H= m + kx , + kx = 2 m 2 2 m

∂T = m˙x, p= ∂ x˙ p ∂H = = x˙ , ∂p m

∂H = kx = −p˙ ∂x

implies (¨x, p¨ ) = −

k (x, p). m

(b) ∂T ∂T = mr2 θ˙ , = m˙r, pθ = ∂ r˙ ∂ θ˙     p2θ 1 1 1 2 − p + , + mµ H= 2m r r2 2a r pr =

p2 ∂H mµ = − θ 3 + 2 = −p˙ r , ∂r mr r

∂H pr = = r˙, ∂pr m

∂H pθ = 2 = θ˙ , ∂pθ mr

∂H = 0 = −p˙ θ . ∂θ Thus, r¨ − rθ˙ 2 = − rµ2 and

d 2˙ dt (r θ ) = 0.

8. (c)  2 2 xˆ , pˆ x = xˆ 2 pˆ 2x − pˆ 2x xˆ 2         = xˆ xˆ , pˆ x xˆ + xˆ pˆ x xˆ , pˆ x + pˆ x xˆ , pˆ x xˆ + xˆ , pˆ x = 4i
ˆxpˆ x + 2
2 . (d) 

         pˆ x , xˆ 2 = xˆ pˆ x , xˆ + pˆ x , xˆ xˆ = −ˆx xˆ , pˆ x − xˆ , pˆ x xˆ = −2i
ˆx.

558

Hints and Answers to Selected Exercises

    10. Lˆ 2 , Lˆ x = Lˆ 2x + Lˆ 2y + Lˆ 2z , Lˆ x . 11. Use (7.267ab). 12. Use (7.306). 15.        Lˆ x , xˆ = yˆ pˆ z − zˆ pˆ y , xˆ = yˆ pˆ z , xˆ − zˆ pˆ y , xˆ ,     Lˆ x , Lˆ y = yˆ pˆ z − zˆ pˆ y , zˆ pˆ x − xˆ pˆ z     = yˆ pˆ x pˆ x , zˆ + xˆ pˆ y zˆ , pˆ z = i
(ˆxpˆ y − yˆ pˆ x ) = i
Lˆ z . 

19.       1 1 1
ω + n2 +
ω + n3 +
ω. EN = En1 + En2 + En3 = n1 + 2 2 2 21.

 ∞ ) ∞    ψ|ψ = ψn |ψψn  ψk |ψψk  (

n=1

=

k=1

∞  ∞ 

ψn |ψ∗ ψk |ψψn |ψk .

n=1 k=1

22. (a) Use the fact that Aˆ is the difference of two Hermitian operators, and then show that ψ|Aˆ ϕ = Aˆ ψ|ϕ for any ψ and ψ. ˆ B ˆ since Aˆ is a linear operator and B ˆ = B ˆ A, ˆ is (b) Use the fact that A a scalar. (c) Aˆ ψ|Aˆ ψ = ψ|(Aˆ )2 ψ = ψ|[Aˆ − A]2 ψ. 24. Note that 

- , ,    pˆ 2 pˆ 2 ˆ = xˆ pˆ x , x + V = xˆ pˆ x , x + xˆ pˆ x , V , xˆ pˆ x , H 2m 2m

and then xˆ pˆ x − pˆ x xˆ = i
. 28. Write |ϕ = aˆ |ψ. Then       1  1    ˆ −
ω ψ = E −
ω ϕ|ϕ = ψ|ˆa |ˆa|ψ = ψ H  2  2  ∗

if |ψ is normalized.

7.12 Exercises

559

By positive definiteness of the inner product, E ≥ 12
ω, and E = 12
ω implies a|ψ = 0. ˆ Lˆ k ] = +ci
(pˆ × α) 29. Show that [H, ˆ k = 0, k = 1, 2, 3. 30.   E − mc2 a1 − cp1 a3 − c(p1 − ip2 )a4 = 0,   E − mc2 a2 − c(p1 + ip2 )a3 + cp3 a4 = 0,   E + mc2 a3 − cp3 a1 − c(p1 − ip2 )a2 = 0,   E + mc2 a4 − c(p1 + ip2 )a1 + cp3 a2 = 0. Existence of a nontrivial solution of this system of equations implies that the determinant of the coefficient matrix is zero, which gives (E2 − m2 c4 − c2 p2 )2 = 0. 33. ψ3 (x) = B sin kx + C cos kx, where B and C are arbitrary constants, is the general solution of ψxx + k2 ψ = 0, k2 = 2mE
2 . For matching the solutions at x = ±a, ψ3 (a) = B sin ak + C cos ak = 0 = ψ2 (a) and ψ3 (−a) = −B sin ak + C cos ak = 0 = ψ1 (−a). For nontrivial solution of B and C, sin ak cos ak = 0. For an even solution, cos ak2 =2 02 and hence π
ak = (2n+1) π2 , which gives the energy levels, En = (2n+1) . Similarly, 8ma2 2 2 2 n π
for an odd solution, ak = nπ and En = 2ma2 . 34. For the free Dirac equation, (cα · 9 p + βmc2 )ψ = Eψ. Or (cαz pz + 2 βmc )ψ = Eψ. Hence,  2  mc 1 cpz σ3 ψ = Eψ, cpz σ3 −mc2 1 which can be explicitly written in the form 

mc2 − E 0   cpz 0

0 2 mc − E 0 −cpz

  ψ1 0 −cpz   ψ2     = 0. 0 ψ3 −mc2 − E ψ4

cpz 0 −mc2 − E 0

2 2 2 2 4 2 This leads to (E − c pz − m c ) = 0, and hence E = E+ , E+ , E− , E− = ± c2 p2z + m2 c4 . 2

35. (a) (pµ pµ )ψ = ( Ec2 − p2 )ψ = m2 c2 ψ. (b) ∂ p → i
∂ = i
= ∂xµ µ

µ



 i
∂ , −i
∇ , c ∂t

560

Hints and Answers to Selected Exercises

  ∂ i
∂ , i
∇ , pµ → i
∂µ = i
µ = ∂x c ∂t   i
∂µ i
∂ µ ψ = m2 c2 ψ. Or,

 −


2

 1 ∂2 2 − ∇ ψ = m2 c2 ψ. c2 ∂t 2

This leads to the KG equation. (c) Plane wave solutions of the form   ψ(r, t) = exp i(k · r − ωt) . They are eigenfunctions of i
∂t∂ and −i
∇ with eigenvalues E =
ω and p =
k respectively. Substituting in the KG equation gives
ω = ± (
ck)2 + m2 c4 (positive and negative energies). This is a consequence of the fact that the KG equation, unlike the Schrödinger equation, is a second-order equation in the time derivative. (d) Both ψ and ψ satisfy the KG equation   1 ∂2 m2 c2 2 ∇ − 2 2 ψ = 2 ψ. c ∂t
Multiplying ψ equation by ψ and ψ equation by ψ, and subtracting these results gives the desired answer. Substituting the plane wave solutions in equation (d) reveals that ρ(r, t) = (
ω/mc2 ) which is not positive definite since (
ω) can be positive as well as negative. Unlike the case of the Schrödinger equation, ρ(r, t) in the present case cannot be thought of as a probability density, which has to be strictly nonnegative. For this reason, the Klein–Gordon equation was abandoned for a long time as inadequate in quantum mechanics.

8.6 Exercises 2. Consider

0

−∞ . . . da

and

∞ 0

5. Use the polarization identity. 6. Use Theorem 3.4.14. 7. Use Theorem 5.11.21. 10. See Debnath (2002).

. . . da separately.

9.11 Exercises

561

11. See Debnath (2002). 12. See Debnath (2002). 13. See Debnath (2002). 14. See Newland (1993, 1994) or Debnath (2002). 15. See Newland (1993, 1994) or Debnath (2002).

9.11 Exercises 2. Use the fact that u − Tx, u − Tx = u, u − 2u, Tx + Tx, Tx $ % = u, u − 2T ∗ u, x + x, T ∗ Tx , and then use F(x) = −2T ∗ u + 2T ∗ Tx. 5. (b) y(x) = 12 e−x + 12 (1 + e)xe−x − 12 . x (d) y0 (x) = sin sin a . 1 8. I(y, z) = x1 1 + (y )2 + (z )2 dx, where Q(x1 , x2 , x3 ) lies on the sphere. 1 13. Maximize A subject to the condition 0 1 + (y )2 dx = L. In other 1  words, maximize the functional I1 (y) = 0 y(x) + 1 + (y )2 − L dx. Answer: (x − α)2 + (y − β)2 = λ2 where α, β and λ are constants.

24. This polynomial is the real part of the binomial expansion of (cos θ + i sin θ )n , where x = cos θ . 30. Use repeated integration by parts to show that  1 ; n <  Dn x2 − 1 xm dx = 0, m = 0, 1, . . . , n − 1, −1

  and then find the leading coefficient of Dn (x2 − 1)n . 31. Use Rodrigues’ formula and the binomial expansion of (x2 − 1)n . 32. (b) n  n−1  2  
 Dn+1 x2 − 1 = Dn D x2 − 1 x −1 n−1    = 2nDn x x2 − 1
n−1  n−1    = 2n xDn x2 − 1 + nDn−1 x2 − 1 and then use Rodrigues’ formula.

562

Hints and Answers to Selected Exercises

34. (a) Use the recurrence relation 32(a) and 32(b). (b) Use 32(b) and 34(a). 35. (a) Multiply the equality in 34(a) by x and subtract from the equality in 32(b). (b) Square and add the equalities in 32(b) and 34(b). 38. Transform into polar coordinates. 41. w = 0 unless λ = n2 . If λ = n2 , wn = An sin nx. 42. (a) w = A sin nx, A[−n2 − (λ − A2 )] = 0. (b) w = A sin nx, A(λ − A2 n2 ) = 0. 44. The term within the first bracket of the equation can be replaced  by a λ 2 2 2 constant µ. Answer: µ = n π , w = An sin nπ x, An = 4 n2 π 2 − 1 . 46. Note that the equality in the square bracket is a constant and can be replaced by a constant a. Then a = an = n2 π 2 , u = A sin nπ x, λ − |A|2 = n2 π 2 . Draw the bifurcation diagram. 47. Square both sides of the equation and integrate from 0 to 1. 52. Maximize the functional  x2 
 I(y) = y(x) + λ y − 1 + (y )2 dx. x1

Answer: (y(x) − b)2 + (x + a)2 = λ2 , where a, b, and λ are constants. 53. In terms of cylindrical coordinates (x, y, z) = (r cos θ, r sin θ, z), points lying on f (x, y, z, ) = x2 + y2 − R2 = 0. On the surface r = R and x˙ (t) = −R sin θ θ˙ and y˙ = R cos θ θ˙ . The problem is to minimize the functional  t2   2 2 1/2 I(x, y, z) = dt R sin θ + cos2 θ θ˙ 2 + z2  =

t1 t2



R2 θ˙ 2 + z2 dt.

t1

The Euler–Lagrange equations are , θ˙ d = 0 and dt R2 θ˙ 2 + z2 These give spiral.

dz dθ

=

z˙ θ˙

, z˙ d = 0. dt R2 θ˙ 2 + z2

= constant = c. Consequently, z = aθ + b, which is a

55. (a) In one case, there are two saddle-node bifurcation points, and in the other case, there are no bifurcation points. 56. (a) The normal form x˙ = λx − x3 has a single equilibrium at x = 0 for √ λ < 0, and three equilibrium points at x = 0, ± λ when λ > 0. The

9.11 Exercises

563

bifurcation at x = 0 and λ = 0 is called a supercritical bifurcation. The normal form x˙ = λx − x3 produces two equilibrium points at the bifurcation point which are unstable. This is called subcritical bifurcation. (b) This normal system has equilibrium points at x = 0 and x = λ. When λ = 0, x˙ = −x2 gives the nonhyperbolic equilibrium point. When λ = 0, there are two equilibrium points and hence, * d˙x −λ = λ − 2x = λ dx

at x = λ at x = 0.

Consequently, x = λ is stable for λ > 0, and unstable for λ < 0; and x = −λ is stable for λ < 0 and unstable for λ > 0. This is called a transcritical bifurcation. It follows from the bifurcation diagram that, at the bifurcation point, the two equilibrium solutions pass through each other and exchange stabilities. 57. (a) For λ < 0, there are no equilibrium points,√while for λ > 0, there are two hyperbolic equilibrium points at x = ± λ. When λ = 0, X(x) = −x2 and x = 0 is a nonhyperbolic equilibrium point. Draw a bifurcation diagram for equilibrium solutions as a function of λ and find stable and unstable equilibria. √ (b) Equilibrium points at x = −λ ± λ + λ2 , and hence no real equilibrium points for −1 < λ < 0 and two equilibrium points otherwise. There are saddle-node bifurcation at λ = 0, x = 0 and λ = −1, x = 1. At equilibrium points d˙x = −2λ − 2x = + λ + λ2 . dx For large x, equilibrium point is stable while the other is unstable. For x % 1 and λ % 1, x˙ = λ − 2λx − x2 ≈ λ − x2 , which is the same as 57 (a). Without any assumption about how large λ is compared to x2 , we examine the neighborhood of the other bifurcation point. Shift the origin using λ = −1 + aµ, x = 1 + by where a and b are constants to be determined later. In terms of y and µ, y˙ = − ab µ − 2aµy − by2 ≈ − ab µ − by2 for µ % 1 and y % 1. With a = −1 and b = 1, this gives the normal form of the saddle-node bifurcation. Thus, λ = −(1 + µ) indicates that the sense of the bifurcation is reversed with respect to λ. 58. (a) The origin is the only equilibrium point and the linearized system is x˙ = −x, y˙ = 0. The eigenvalues are 0 and −1 and so this is a nonhyperbolic equilibrium point. We construct a Lyapunov function by trying V = x2 + ay2 which is positive definite for a > 0 and V(0, 0) = 0. Fur2 2 4 ˙ ther, dV dt = −2x + 2(a − 2)xy − 2ay . With a = 2, V < 0 for all x and y except (x, y) = (0, 0). This leads to the solution.

564

Hints and Answers to Selected Exercises

(b) x¨ + x˙ + x + x2 = 0. Consider V = 12 (x2 + y2 ) + 13 x3 . V(0, 0) = 0 and V is positive in the region D = {(x, y)|y2 > −(x2 + 23 x3 )}. Consider the curve V = 16 which passes through the saddle point (−1, 0) and the point ( 12 , 0). If V = V0 < 16 , there is a curve that encloses the origin, but not the saddle point. For the domain of attraction of the equilibrium point at (0, 0), consider a domain D0 which is a subset of D given by 2 V < V0 < 16 with V0 close to 16 . Since dV dt = −y ≤ 0, it follows that the origin is Lyapunov stable. 59. (a) (0, 0) is an equilibrium point of this nonlinear system. At (0, 0), the linearized system is x˙ = y˙ = 0. Both eigenvalues are zero. Consider a Lyapunov function in the form V = axy2 − x3 , where V(0, 0) = 0. 2 2 4 4 4 4 ˙ Since dV dt = 3(1 − a)x y − ay − 3x . With a = 1, V = −y − 3x < 0, 2 2 V = x(y − x ) = 0 when x = 0 or y = ± x so that V changes sign six times on any circle with center at the origin. In every neighborhood of (0, 0), there is at least one point where V and V˙ have the same sign. Thus, the origin is unstable. (b) The Jacobian at the equilibrium point, (0, 0) is   2 0 J= , 1 −1 which has eigenvalues 2 and −1 with the corresponding eigenvectors (3, 1)T and (0, 1)T . Hence, the local stable manifold is the y-axis and the local unstable manifold is the line 3y = x. The solution is ,  2a 4  −t −t t x(t) = , y(t) = e b + 2e − log 2 + ae , a + 2e−t a where a = 0 and b are constants. There is also the obvious solution x = 0, y = be−t , y-axis, which yields the local stable manifold and the global stable manifold. 61. There are three equilibrium points, one at x = y = z = 0 and two at √ x = 27, y = z = ±6 2, each of which is unstable. The two equilibrium points away from the origin each have a two-dimensional unstable manifold, associated with complex eigenvalues, and thus oscillatory nature, and a one-dimensional stable manifold. The Lorenz system has very interesting dynamics. Solutions bounce back and forth between the equilibrium points away from the origin, continually being attracted toward an equilibrium point along a trajectory close to the stable manifold and then spiraling away close to the unstable manifold for the solution with x(0) = y(0) = z(0) = 1.

Bibliography

The following bibliography is not, by any means, a complete one for the subject. For the most part, it consists of books and papers to which reference is made in the text. Many other selected books and papers related to material in this book have been included so that they may serve to stimulate new interest in future study and research. Adams, R.A., and Fournier, P., Sobolev Spaces, Pure and Applied Mathematics, Vol. 65 (Second Edition), Academic Press, 2003. Antosik, P., and Swartz, C., Matrix Methods in Analysis, Springer-Verlag, Berlin, 1985. Arrowsmith, D.K., and Place, C.M., An Introduction to Dynamical Systems, Cambridge University Press, Cambridge, 1992. Balakrishnan, A.V., Applied Functional Analysis, Springer-Verlag, New York, 1976. Balakrishnan, A.V., Introduction to Optimization Theory in a Hilbert Space, Springer-Verlag, New York, 1971. Banach, S., Théorie des opérations linéaires, Chelsea, New York, 1955. Battle, G., A block spin construction of ondelettes, Part I: Lemarié functions, Commun. Math. Phys. 110 (1987), 601–615. Berkovitz, L., Optimal Control Theory, Springer-Verlag, New York, 1975. Beylkin, G., On the representation of operators in bases of compactly supported wavelets, SIAM J. Numer. Anal. 29 (1992), 1716–1740. Beylkin, G., Coifman, R., and Rokhlin, V., Fast wavelet transforms and numerical algorithms, Comm. Pure Appl. Math. 44 (1991), 141–183. Cheney, E.W., Introduction to Approximation Theory, McGraw-Hill, New York, 1966.

565

566

Bibliography

Chui, C.K., An Introduction to Wavelets, Academic Press, Boston, 1992. Chui, C.K., and Wang, J.Z., A cardinal spline approach to wavelets, Proc. Amer. Math. Soc. 113 (1991), 785–793. Chui, C.K., and Wang, J.Z., On compactly supported spline wavelets and a duality principle, Trans. Amer. Math. Soc. 330 (1992), 903–915. Coddington, E.A., and Levinson, N., Theory of Ordinary Differential Equations, McGraw-Hill, New York, 1955. Cohen, A., Biorthogonal Wavelets, in Wavelets — A Tutorial in Theory and Applications, Academic Press (1992b), 123–152. Cohen, A., Daubechies, I., and Feauveau, J., Biorthogonal basis of compactly supported wavelets, Comm. Pure and Appl. Math. 45 (1992a), 485–560. Coifman, R.R., Jones, P., and Semmes, S., Two elementary proofs of the L2 boundedness of Cauchy integrals on Lipschitz curves, J. Amer. Math. Soc. 2 (1989), 553–564. Coifman, R.R., Meyer, Y., and Wickerhauser, M.V., Wavelet analysis and signal processing, in: Wavelets and Their Applications (Eds. Ruskai et al.) (1992a), 153–178; and Size Properties of Wavelet Packets (1992b), 453–470, Jones and Bartlett, Boston. Curtain, R.F., and Pritchard, A.J., Functional Analysis in Modern Applied Mathematics, Academic Press, New York, 1977. Daubechies, I., Ten Lectures on Wavelets, CBMS-NSF Regional Conference Series in Applied Math., Vol. 61, SIAM, Philadelphia, PA, 1992. Daubechies, I., The wavelet transform, time-frequency localization and signal analysis, IEEE Trans. Inform. Theory 36 (1990), 961–1005. Daubechies, I., Orthogonal bases of compactly supported wavelets, Commun. Pure Appl. Math. 41 (1988), 909–996. Daubechies, I., Grossmann, A., and Meyer, Y., Painless nonorthogonal expansions, J. Math. Phys. 27 (1986), 1271–1283. Debnath, L., Integral Transforms and Their Applications (Second Edition), CRC Press, Boca Raton, FL, 2005. Debnath, L., Nonlinear Partial Differential Equations for Scientists and Engineers (Second Edition), Birkhäuser, Boston, 2005. Debnath, L., Wavelet Transforms and Their Applications, Birkhäuser, Boston, 2002. De Boor, C., Approximation Theory, Proceedings of Symposia in Applied Mathematics, Vol. 36, American Mathematical Society, Providence, RI, 1986. Dieudonné, J., Foundations of Modern Analysis, Academic Press, New York, 1969. Dirac, P.A.M., The Principles of Quantum Mechanics (Fourth Edition), Oxford University Press, Oxford, 1958. Duffin, R.J., and Schaeffer, A.C., A class of nonharmonic Fourier series, Trans. Amer. Math. Soc. 72 (1952), 341–366.

Bibliography

567

Dunford, N., and Schwartz, J.T., Linear Operators, Part I, General Theory, Interscience, New York, 1958. Friedman, A., Foundations of Modern Analysis, Dover, New York, 1982. Gabor, D., Theory of communications, J. Inst. Electr. Eng. London 93 (1946), 429–457. Garabedian, P.R., Partial Differential Equations, Wiley, New York, 1964. Glimm, J., and Jaffe, A., Quantum Physics (Second Edition), Springer-Verlag, New York, 1987. Gould, S.H., Variational Methods for Eigenvalue Problems, Toronto University Press, Toronto, 1957. Grossmann, A., and Morlet, J., Decomposition of Hardy functions into square integrable wavelets of constant shape, SIAM J. Math. Anal. 15 (1984), 723– 736. Guckenheimer, J., and Holmes, P.J., Nonlinear Oscillations, Dynamical Systems and Bifurcation of Vector Fields, Springer-Verlag, New York, 1983. Halmos, P.R., Measure Theory, Springer-Verlag, New York, 1974. Hilbert, D., Grundzüge einer allgemeinen Theorie der linearen Integralgleichungen, Leipzig, 1912. Holschneider, M., Wavelets, An Analysis Tool, Oxford University Press, Clarendon Press, Oxford, 1995. Hutson, V., and Pym, J.S., Applications of Functional Analysis and Operator Theory, Academic Press, New York, 1980. Iooss, G., and Joseph, D.D., Elementary Stability and Bifurcation Theory, Springer-Verlag, New York, 1981. Jacuch, J.M., Foundations of Quantum Mechanics, Addison-Wesley, Reading, MA, 1968. Jones, D.S., Generalized Functions, McGraw-Hill, New York, 1966. Kantorovich, L.V., and Akilov, G.P., Functional Analysis in Normed Spaces, Pergamon Press, London, 1964. Kanwal, R.P., Generalized Functions (Second Edition), Birkhäuser, 1998. Keller, J.B., and Antman, S., Bifurcation Theory and Nonlinear Eigenvalue Problems, W.A. Benjamin, New York, 1969. Kolmogorov, A.N., and Fomin, S.V., Elements of the Theory of Functions and Functional Analysis, Vol. 1, Graylock Press, Rochester, New York, 1957; Vol. 2, Graylock Press, Albany, New York, 1961. Kolmogorov, A.N., and Fomin, S.V., Introductory Real Analysis, Prentice-Hall, New York, 1970. Kreyn, S.G., Functional Analysis, Foreign Technology Division WP-AFB, Ohio, 1967. Kreyszig, E., Introductory Functional Analysis with Applications, Wiley, New York, 1978. Landau, L.D., and Lifshitz, E.M., Quantum Mechanics, Non-relativistic Theory, Pergamon Press, London, 1959.

568

Bibliography

Lax, P.D., and Milgram, A.N., Parabolic equations, contribution to the theory of partial differential equations, Ann. of Math. Stud., No. 33, Princeton, 1954, 167–190. Lemarié, P.G., Bases d’ondelettes sur les groupes de Lie stratifiés, Bull. Soc. Math. France 117 (1989), 211–232. Lemarié, P.G., Ondelettes à localisation exponentielle, J. Math. Pures Appl. 67 (1988), 227–236. Lemarié, P.G., and Meyer, Y., Ondellettes et bases hilbertiennes, Revista Mat. Iberoamer. 2 (1986), 1–18. Lions, J.L., and Stampacchia, G., Variational inequalities, Comp. Pure Appl. Math. 20 (1967), 493–519. Liusternik, L.A., and Sobolev, V.J., Elements of Functional Analysis (Third English Edition), Hindustan Publishing Co., New Delhi, 1974. Lorenz, E.N., Deterministic non-periodic flows, J. Atmos. Sci. 20 (1963), 130– 141. Luenberger, D.G., Optimization by Vector Space Methods, Wiley, New York, 1969. Mackey, G.W., The Mathematical Foundations of Quantum Mechanics, W.A. Benjamin, New York, 1963. MacNeille, H.M., A unified theory of integration, Proc. Nat. Acad. Sci. USA 27 (1941), 71–76. Mallat, S., Multiresolution approximations and wavelet orthonormal basis of L2 (R), Trans. Amer. Math. Soc. 315 (1989a), 69–88. Mallat, S., A theory for multiresolution signal decomposition: The wavelet representation, IEEE Trans. Pattern Anal. Machine Intell. 11 (1989b), 678–693. Mallat, S., Multiresolution representation and wavelets, Ph.D. Thesis, University of Pennsylvania, 1988. Merzbacher, E., Quantum Mechanics (Second Edition), Wiley, New York, 1961. Meyer, Y., Ondelettes, fonctions splines et analyses graduées, Lecture Notes, University of Torino, Italy, 1986. Meyer, Y., Wavelets and Operators, Cambridge University Press, Cambridge, 1992. Meyer, Y., Wavelets, Algorithms and Applications, SIAM, Philadelphia, PA, 1993. ´ Mikusinski, J., A theorem on vector matrices and its applications in measure theory and functional analysis, Bull. Acad. Polon. Sci. Sér. Sci. Math. Astronom. Phys. 18 (1970), 193–196. ´ Mikusinski, J., Bochner Integral, Birkhäuser, Basel, 1978. ´ Mikusinski, P., and Taylor, M.D., An Introduction to Multivariable Analysis: From Vector to Manifold, Birkhäuser, 2001. Myint-U, T., and Debnath, L., Partial Differential Equations for Scientists and Engineers (Fourth Edition), Birkhäuser, Boston, 2005.

Bibliography

569

Neumann, J.V., Mathematical Foundations of Quantum Mechanics, Princeton University Press, Princeton, 1955. Newland, D.E., Harmonic wavelet analysis, Proc. Roy. Soc. London Ser. A 443 (1993), 203–225. Newland, D.E., Harmonic and musical wavelet, Proc. Roy. Soc. London Ser. A 444 (1994), 605–620. Reed, M., and Simon, B., Methods of Modern Mathematical Physics, Vol. 1, Functional Analysis, Academic Press, New York, 1972. Riesz, F., and Sz-Nagy, B., Functional Analysis (Second Edition), Frederick Ungar, New York, 1955. Rivlin, T.J., An Introduction to the Approximation of Functions, Dover, New York, 1969. Roach, G.F., Green’s Functions (Second Edition), Cambridge University Press, Cambridge, 1982. Sattinger, D.H., Topics in Stability and Bifurcation Theory, Lecture Notes in Mathematics, Vol. 309, Springer-Verlag, New York, 1973. Schechter, M., Modern Methods in Partial Differential Equations, McGraw-Hill, New York, 1977. Schwartz, L., Théorie des distributions, Vols. I and II, Herman and Cie, Paris, 1950, 1951. Shilov, G.E., Generalized Functions and Partial Differential Equations, Gordon and Breach, New York, 1968. Sobolev, S.L., Partial Differential Equations of Mathematical Physics, Pergamon Press, London, 1964. Stakgold, I., Boundary Value Problems of Mathematical Physics, Macmillan, New York, 1968. Swartz, C., An Introduction to Functional Analysis, Marcel Dekker, New York, 1992. Taylor, A.E., Introduction to Functional Analysis, Wiley, New York, 1958. Tricomi, F.G., Integral Equations, Interscience, New York, 1957. Walter, G., Wavelets and Other Orthogonal Systems with Applications, CRC Press, Boca Raton, FL, 1994. Yosida, K., Functional Analysis (Fourth Edition), Springer-Verlag, New York, 1974. Young, L.C., Calculus of Variations and Optimal Control Theory, W.B. Saunders Company, Philadelphia, PA, 1969. Zayed, A.I., Advances in Shannon’s Sampling Theory, CRC Press, Boca Raton, FL, 1993. Zemanian, A.H., Distribution Theory and Transform Analysis, McGraw-Hill, New York, 1965. Ziemer, W.P., Weakly Differentiable Functions, Springer-Verlag, New York, 1989.

This page intentionally left blank

Index

B(E1 , E2 ), 28 Aλ , 188 B(x, r), 14 Dα f , 101 H m (), 102 H m (), 302 L1 -norm, 52 L1 (R), 46 L1 (RN ), 78 L2,ρ ([a, b]), 100 L∞ -norm, 77, 79 L∞ (R), 77 L∞ (RN ), 79 Lp -norm, 74 Lp (R), 74 Lp (RN ), 79 N ϕ, 470 N ψ , 470 S(x, r), 14 S⊥ , 127 W m,p (), 300 · ∞ , 77 · p , 74 · 0,p, , 302 u m,∞, , 300 u m,p, , 300 δmn , 106

dim E, 8 ·, ·, 94 C, 2 F, 2 N, 4 R, 2 CN , 3 RN , 3 C(), 4 C ∞ (), 4 C k (), 4 D(RN ), 289 D(L), 25 G(L), 25 I, 146 N (L), 25 P(), 4 R(L), 25 C r -diffeomorphism, 301 ∇ 2 , 307 cl S, 16 ess sup f , 77, 79 lub S, 42 sgn, 117 span A, 8 B(x, r), 14 ρ(A), 188

571

572

Index

σ (A), 188 σ -additive measure, 69
, 309 f ∗ g, 82 f  f1 + f2 + · · ·, 45, 51 fn → f a.e., 56 fn → f i.n., 54 lp , 4 x ⊥ y, 98 w xn → x, 102 L1 (R), 53

A Abel, N.H., 238 Abel’s equation, 237 Abel’s formula, 250 Abel’s problem of tautochronous motion, 498 absolutely convergent series, 22 abstract minimization problem, 508 action, 385 adjoint boundary conditions, 241 adjoint of a densely defined operator, 204 adjoint of a differential operator, 240 adjoint operator, 158 admissibility condition, 437 admissible set, 490 admittance, 272 a.e., 55 Airy function, 343 Airy, G.B., 309 Airy integral, 343 analyzing wavelet, 441 angular momentum, 362 annihilation operator, 409 anomalous Zeeman effect, 416 antiderivative of a distribution, 297 antilinear functional, 134 Antosik, P., 30 Apollonius’ identity, 136 approximate eigenvalue, 195 associated Laguerre equation, 246 associated Laguerre functions, 246 associated Legendre functions, 244 asymptotically stable solution, 526 atomic units, 389

autonomous dynamical system, 526

B Baire, R.-L., 30 Banach fixed point theorem, 33 Banach, S., 19 Banach space, 20 Banach–Steinhaus theorem, 31 band-limited signal, 523 bandwidth of a signal, 523 basic representation of a step function, 40 basis, 8 Bessel equation, 246 Bessel, F.W., 111 Bessel’s equality and inequality, 111 best approximation, 517 bifurcation, 531 bifurcation point, 531 biharmonic equation, 309, 339 biharmonic wave equation, 308 bilinear concomitant, 241 bilinear functional, 151 Bohr, N.H.D., 371 Born, M., 384 Bose, S.N., 419 Bosons, 419 bounded bilinear functional, 152 bounded linear mapping, 27 bounded quadratic form, 152 bounded set, 17 branch point, 531

C Cantor set, 87 Cantor, G., 87 Carleson, L., 127 Cartesian product of vector spaces, 7 catenary, 496 catenoid, 496 Cauchy, A.L., 19 Cauchy sequence, 19 Cayley, A., 212 Cayley transform, 212 chain rule, 485 change of variables formula, 66 Chebyshev operator, 245

Index Chebyshev, P.L., 138 Chebyshev polynomial, 138 Chebyshev polynomials, 245 Chebyshev polynomials of the second kind, 522 classical solution, 304 closed ball, 14 closed operator, 208 closed set, 14 closest point property, 128 closure of a set, 16 coercive bilinear functional, 156 commutator, 368 commuting operators, 149 compact operator, 180 compact set, 17 compatible observables, 369 complementary observables, 369 complementary projection operator, 177 complete normed space, 20 complete orthonormal sequence, 113 completely continuous operator, 180, 185 completion of a normed space, 24 complex vector space, 2 complex-valued step function, 71 conjugate quadratic filters, 459 conjugate variables, 357 conjugate-linear functional, 134 conservative field, 352 conserved function, 353 conserved observable, 399 constant of motion, 353, 404 continuous mapping, 26 continuous wavelet transform, 440 contraction mapping, 32 contraction mapping theorem, 219 control function, 511 control of chemical reaction with nonlinear cost, 516 convergence almost everywhere, 56 convergence in D(), 299 convergence in a normed space, 10 convergence in norm of integrable functions, 54 convergence of distributions, 295 convergence of test functions, 290

573

convergent series, 22 convex function, 489 convex set, 128 convolution, 82 convolution theorem, 263 correspondence principle, 371 creation operator, 409

D d’Alembert, J.L.R., 332 d’Alembert solution, 331 d’Alembertian operator, 309 Daubechies D4 wavelet, 472 Daubechies, I., 433, 434 Daubechies wavelet, 470 Daubechies wavelets and algorithms, 467 de Broglie, L.V.P.R., 382 degenerate eigenvalue, 187 degenerate kernel, 233 degree of degeneracy of an eigenvalue, 187 dense subset, 17 densely defined operator, 203 diagonal theorem, 30 differential operator, 147 diffusion equation, 308, 319, 329 dimension, 8 dimension of the state space, 366 Dirac delta distribution, 293 Dirac equation, 420 Dirac Hamiltonian, 420 Dirac matrices, 421 Dirac, P.A.M., 293 Dirichlet integral, 311 Dirichlet, J.P.G.L., 306 Dirichlet problem, 306, 323 discrete Fourier transform, 456 discrete wavelet coefficients, 445 dispersion relation, 386 distribution, 291 distributional convergence, 295 distributional derivative, 294 distributional solution, 305 distributions on an open subset of RN , 299

574

Index

Ehrenfest, P., 392 Ehrenfest’s theorem, 392, 400 eigenbasis, 370 eigenfunction, 186 eigenspace, 186 eigenvalue, 186 eigenvalue problem, 188 eigenvector, 186 elliptic bilinear functional, 156 equality almost everywhere, 55 equality of operators, 149 equation of continuity for probability flow, 383 equations of motion, 363 equilibrium solution, 526 equivalent Cauchy sequences, 24 equivalent functions, 53 equivalent norms, 11 essential supremum, 77 Euclidean norm, 9 Euclidean space, 98 Euler, L., 237 Euler gamma function, 237 Euler–Lagrange equation, 494 Euler–Lagrange variational principle, 494 expectation value, 373 extension of an operator, 203 extremum, 490

Fermions, 419 filter conjugate, 459 finite dimensional vector space, 8 finite rank operator, 182 finite-dimensional operator, 182 fixed point, 32, 218 formally self-adjoint differential operator, 242 Fourier coefficients, 127 Fourier, J.B.J., 112 Fourier series, 127 Fourier transform in L1 (R), 258 Fourier transform in L2 (R), 264 frame, 445 frame bounds, 445 frame operator, 449 Franklin wavelet, 474 Fréchet differential, 482 Fréchet, M.R., 481 Fredholm alternative, 222 Fredholm, E.I., 218 Fredholm equation, 218 Fredholm integral equation with convolution kernel, 277 Fredholm nonhomogeneous linear integral equation of the second kind, 224 Fredholm operator, 160 free particle de Broglie wave, 387 frequency of a signal, 523 Friedrichs’ first inequality, 325 Fubini, G., 80 Fubini’s theorem, 80 function of an operator, 200 function space, 3 functional, 29 fundamental matrix, 512 fundamental set of solutions, 512 fundamental solution, 306

F

G

father wavelet, 452 Fatou, P.J.L., 61 Fatou’s Lemma, 61 Fejér, L., 124 Fejér’s kernel, 124 Fermat’s principle in optics, 497

Gabor, D., 434 Gabor transform, 434 gate function, 523 Gateaux differentiable, 478 Gateaux differential, 478 Gauss, J.C.F., 262

divergence form of a differential operator, 328 domain of a differential operator, 240 domain of a mapping, 25 domain of dependence, 332 dominated convergence theorem, 60 dual frame, 451 dual space, 30

E

Index Gegenbauer, L.B., 139 Gegenbauer polynomials, 139 general Parseval’s relation, 267 generalized coordinates, 354 generalized force, 357 generalized Fourier coefficients, 112 generalized Fourier series, 112 generalized momentum, 357 generating function of the multiresolution analysis, 456 generator of the infinitesimal unitary transformation, 397 generators of infinitesimal translations, 363 geodesic, 502 gradient, 483 gradient of a functional, 479 Gram, J.P., 110 Gram–Schmidt orthonormalization process, 110 graph of a mapping, 25 Green, G., 253 Green’s first identity, 309 Green’s function, 253, 307 Green’s second identity, 310 Grossmann, A., 434 ground state energy, 409, 410 group velocity, 386

H Haar, A., 270 Haar function, 269 Haar wavelet, 438, 463 Hamilton equations of motion, 357 Hamiltonian, 358, 496 Hamiltonian operator, 380 Hamilton’s canonical equations, 359 Hamilton’s function, 358, 496 Hamilton’s principal function, 385 Hamilton’s principle, 496 Hamilton’s variational principle, 360 Hamilton, W.R., 357 Hammerstein operator, 484 heat equation, 308 Heaviside function, 292 Heaviside, O., 292 Heisenberg equations of motion, 402

575

Heisenberg picture, 395 Heisenberg uncertainty principle, 377 Heisenberg, W.K., 372 Helmholtz equation, 315, 316 Helmholtz’s equation, 308 Hermite, C., 108 Hermite equation, 247 Hermite operator, 247 Hermite polynomials, 108, 247 Hermitian matrix, 159 Hilbert, D., 93 Hilbert–Schmidt theorem, 196 Hilbert space, 99 Hilbert space isomorphism, 121, 268 Hilbert transform, 277 Hölder, O.L., 5 Hölder’s inequality for functions, 74 Hölder’s inequality for sequences, 5 holonomic dynamical system, 355 homogeneous Dirichlet problem, 306 homogeneous Neumann problem, 306 homogeneous Volterra equation, 232

I ideal low-pass filtering, 523 idempotent operator, 176 identity operator, 146 image of a point, 25 image of a set, 25 implicit function theorem, 535 i.n., 54 infinite dimensional vector space, 8 inner product, 94 inner product space, 94 input function, 272 integral equations, 218 integral of a step function, 41 integral operator, 148 integral over an interval, 62 interaction picture, 395 intrinsic angular momentum, 416 inverse Fourier transform in L2 (R), 267 inverse image of a set, 25 inverse operator, 163 inversion formula for wavelet transforms, 442 invertible operator, 163

576

Index

kernel, 218, 436 kernel of an integral operator, 148 Klein–Gordon equation, 309, 320 Korteweg–de Vries equation, 342 Korteweg, D.J., 342 Kronecker delta, 106 Kronecker, L., 106

Legendre polynomials, 107, 244 Leibniz formula, 299 Levi, B., 60 linear combination, 7 linear dependence, 7 linear independence, 7 linear mapping, 25 linear momentum, 352 linear operator, 145 linear transformation, 145 linear-quadratic control problem, 511 Lions–Stampacchia theorem, 509 Liouville, J., 247 local extremum, 490 locally integrable function, 62, 73, 79 Lorenz equation, 544 low-pass filter, 456 low-pass filter function, 523 lowest state energy, 409, 410 Lyapunov, A.M., 528 Lyapunov’s theorem, 528

L

M

Lagrange equation of motion, 354 Lagrange equations of motion, 356, 496 Lagrange, J.-L., 250 Lagrange’s identity, 250 Lagrangian, 496 Lagrangian function, 353 Laguerre differential equation, 245 Laguerre, E.N., 138 Laguerre operator, 245 Laguerre polynomials, 138, 246 Laplace equation, 307 Laplace operator, 307 Laplace, P.-S., 307 Laplace’s equation in a half-plane, 334 law of conservation of energy, 359 Lax–Milgram theorem, 157 least-squares approximation, 518 Lebesgue dominated convergence theorem, 60 Lebesgue, H.L., 39 Lebesgue integrable function, 45, 72, 78 Lebesgue measure, 68 Legendre, A.-M., 107 Legendre operator, 244

MacNeille, M., 39 Malvar, H., 434 Malvar wavelets, 434 mathematical representation, 367 matrix Riccati equation, 514 Maxwell, J.C., 364 mean value theorem, 481 measurable function, 70, 79 measurable set, 68, 79 measure, 68 measurement, 368 measurement of a classical observable, 367 metric in a normed space, 13 Mexican hat wavelet, 443 Meyer, Y., 434, 436 ´ Mikusinski, J., 30, 39 minimum arc length, 494 Minkowski, H., 5 Minkowski’s inequality for functions, 75 Minkowski’s inequality for sequences, 6 mixed problem, 307 modulated Gaussian function, 261 momentum wave function, 389

isometric operator, 166 isomorphic Hilbert spaces, 121 isothermal curves, 335

J Jacobian matrix, 480 Jacobi, C.G.J., 139 Jacobi differential equation, 245 Jacobi operator, 245 Jacobi polynomials, 139, 245 Jacobi’s identity, 424

K

Index monotone almost everywhere, 60 monotone convergence theorem, 60 Morlet, J., 433 mother wavelet, 441, 455 motion of a wave-packet in a field of force, 391 multiple eigenvalue, 187 multiplication of distributions by smooth functions, 299 multiplication operator, 148 multiplicity of an eigenvalue, 186 multiplier, 148 multiresolution, 435 multiresolution analysis, 434, 452

N Navier, C.L.M.H., 337 Navier–Stokes equation, 337 Neumann boundary value, 325 Neumann, C.G., 190 Neumann problem, 306 Neumann series, 190 Newland wavelet, 475 Newton, I., 352 Newton’s second law of motion, 352 noncommuting operators, 149 nondegenerate eigenvalue, 186 nonholonomic dynamical system, 355 nonlinear Fredholm integral equation, 225 nonseparable Hilbert space, 119 norm, 8, 260, 263 norm in an inner product space, 97 norm of a bilinear functional, 152 norm of a quadratic form, 152 norm of uniform convergence, 11 normal operator, 165 normed space, 9 null function, 53 null operator, 146 null set, 55 null space, 25

O observable, 365 observables, 363, 367 observation of a classical observable, 367

577

octaves, 435 one-sided shift operator, 166 open ball, 14 open set, 14 operator norm, 146 optimal control, 511 optimal error, 517 optimal solution, 517 optimal trajectory, 511 orbital quantum number, 394 orthogonal complement, 127 orthogonal decomposition, 131 orthogonal projection, 131 orthogonal projection operator, 175 orthogonal projection operators, 177 orthogonal sets, 127 orthogonal state vectors, 366 orthogonal system, 105 orthogonal vectors, 98 orthonormal basis, 114 orthonormal sequence, 106 orthonormal system, 105 orthonormal wavelet, 463 output function, 272

P parallelogram law, 97 Parseval, M.-A., 115 Parseval’s formula, 115 Parseval’s relation, 265 Parseval’s relation for wavelet transforms, 441 Pauli spin matrices, 415 Pauli, W.E., 415 periodic boundary conditions, 240 periodic Sturm–Liouville system, 248 Picard, 221 Picard theorem, 221 Plancherel, M., 268 Plancherel theorem, 268 Planck constant, 364 Planck, M., 364 Planck’s simple harmonic oscillator, 411 pointwise convergence, 11 Poisson bracket, 362 Poisson equation, 307, 313 Poisson integral, 334

578

Index

Poisson, S.D., 307 polarization identity, 136, 152 positive bilinear functional, 152 positive definite operator, 175 positive operator, 169 Postulate I of quantum mechanics, 365 Postulate II of quantum mechanics, 367 Postulate III of quantum mechanics, 371 Postulate IV of quantum mechanics, 372 Postulate V of quantum mechanics, 376 Postulate VI of quantum mechanics, 380 pre-Hilbert space, 94 principal quantum number, 394 principle of linearized stability, 527 principle of quantization, 372 principle of superposition, 315 probability current density, 383 probability density, 383 probability flux, 383 product of operators, 149 projection operator, 175 proper subspace, 4 Pythagorean formula, 98, 110

Q quadratic form, 152 quantum number, 410

R Rademacher functions, 117 Rademacher, H., 117 range of a mapping, 25 real vector space, 2 regular distribution, 292 regular Sturm–Liouville system, 247 regular values, 188 relative extremum, 490 relative maximum, 490 relative minimum, 490 representer of a functional, 134 resolution of a pulse, 273 resolvent, 188 resolvent kernel, 228 resolvent set, 188 Riemann function, 319 Riemann, G.F.B., 64 Riemann–Lebesgue Lemma, 260

Riesz, F., 59 Riesz representation theorem, 133 Riesz’s lemma, 18 Riesz’s lemma for spectral factorization, 470 Robin problem, 307 root-mean-square deviation, 373 Rutherford, E., 364

S sampling integral representation, 525 scalar, 2 scaling function, 452 Schmidt, E., 110 Schrödinger equation, 380, 392 Schrödinger, E.R.J.A., 309, 380 Schrödinger picture, 395 Schwartz, L., 287 Schwartz distribution, 287, 291 Schwarz, H.A., 96 Schwarz’s inequality, 96 second Fréchet differential, 485 second-order elliptic equations, 327 self-adjoint differential operator, 242 self-adjoint operator, 159, 206 semiopen interval in RN , 78 separable Hilbert space, 119 separable kernel, 233 separated boundary conditions, 239 sequence space, 4 sesquilinear functional, 151 set of measure zero, 55 Shannon, C.E., 270 Shannon function, 270 Shannon sampling function, 523 Shannon sampling theorem, 525 Shannon scaling function, 523 Shannon wavelet, 466 short-time Fourier transform, 434 simple eigenvalue, 186 singular distribution, 292 singular Sturm–Liouville system, 247 smooth function, 289 Snell law of refraction of light, 497 Sobolev, S.L., 102 Sobolev space W m,p (), 300 Sobolev spaces, 101

Index spanned space, 8 spectral radius, 188 spectral theorem for compact selfadjoint operators, 198 spectrum, 188 spectrum of a signal, 523 sphere, 14 spherical harmonics, 394 spin of a particle, 416 spinor, 422 spinor wave function, 422 square root of a positive operator, 173 stability criterion, 528 stable solution, 526 standard deviation, 377 state equation, 516 state function, 367, 511 state of a physical system, 365 state of a quantum system, 365 state transition matrix, 512 states, 363 stationary point, 491 stationary state, 380 Steinhaus, H.D., 30 step function, 40, 78 Stokes, G.G., 337 strictly convex function, 489 strictly positive bilinear functional, 152 strictly positive operator, 175 strong convergence, 38 strong convergence in an inner product space, 102 Sturm, J.C.F., 247 Sturm–Liouville system, 247 subspace of a vector space, 4 successive approximations, 219 summability kernel, 123 support of a function, 41 Swartz, C., 30 symmetric bilinear functional, 152 symmetric operator, 206 synthesis of a pulse, 273

T telegrapher’s equation, 308 tent function, 86 test function, 289

579

test functions on an open subset of RN , 298 tight frame, 445 time-dependent Schrödinger equation, 380 time-dependent state vector, 365 time-development operator, 397 time-domain convolution, 523 time-evolution operator, 397 time-invariant control problem, 511 total angular momentum, 419 total time derivative, 400 transfer function, 274 transfer function of a filter, 524 triangle inequality, 9, 97 trigonometric Fourier series, 122 two-sided shift operator, 212

U unbounded operator, 202 uncertainty, 377 uncertainty principle, 377 uniform convergence, 10, 28 uniformly elliptic differential operator, 328 unit impulse, 272 unitary operator, 167 unitary space, 94 universal constant, 364 unstable solution, 526

V value of the observable, 367 variational formulation of a problem, 324 variational inequality problem, 508 variational problems, 477 vector, 2 vector space, 2 vibrating string, 504 virial theorem, 428 Volterra equation of the first kind, 218 Volterra equation of the second kind, 218, 229 Volterra, V., 218 von Helmholtz, H.L.F., 308 von Neumann, J., 93

580

Index

vorticity transport equation, 338

W Walsh functions, 118 Walsh, J., 118 wave equation, 308, 318, 331 wave function, 367 wave–particle duality, 387 wavelet, 437 weak convergence in an inner product space, 102 weak distributional convergence, 295 weak formulation of a problem, 324 weak Parseval’s relation, 265

weak solution, 304, 326 weight function, 244, 518 weighted average, 375 window-Fourier transform, 434 Wronskian, 250 Wronski, J.H., 250

Y Young’s inequality, 90 Young, W.H., 90

Z zero vector, 2